Blog

Healthcare recruiter and international nurse professional reviewing credentials and employment documents at office table

In this article, we'll go through the current list of Cortex AI functions and learn how to navigate between them. For starters, let's create a system to organize them. We'll start with the functions that take text-only input, classifying them into row-level and aggregation functions, and then move on to the ones that accept multiple formats.<figure id="figure-e6b9m6dcei9mu2thhix"><img id="e6b9m6dcei9mu2thhix" src="https://cdn.durable.co/blocks/31IYMMyQQOgPhUiI4IhjmE0mqPOC785O3YXGInkhQWoMugyPkHBlGMYgBMsr0Jk7.png" alt="" title="" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="List of Cortex AI Functions in Snowsight" data-positionx="undefined" data-positiony="undefined" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/31IYMMyQQOgPhUiI4IhjmE0mqPOC785O3YXGInkhQWoMugyPkHBlGMYgBMsr0Jk7.png" data-mce-style="aspect-ratio: 16/9; object-position: center center;"><figcaption class="body-normal text-gray-500 mt-2">List of Cortex AI Functions in Snowsight</figcaption></figure><h2>Functions that accept only text as input</h2>Row-level text functionsAs you can see below, all four functions perform one specific action. Each takes its own set of arguments in addition to the text value.AI_TRANSLATE - an updated version of SNOWFLAKE.CORTEX.TRANSLATE that currently supports 23 languages, and can also handle text that mixes several of them.AI_SENTIMENT - extracts sentiment from the text, with an option to specify which categories to evaluate.AI_REDACT - detects and redacts PII. You can pick which of the 13 supported categories to redact, or run it in detect mode to locate PII without redacting it.SNOWFLAKE.CORTEX.SUMMARIZE - summarizes the text. It is limited by the size of the model's context window.Here is an example of using AI_REDACT:SELECT &nbsp; &nbsp; original_text, &nbsp; &nbsp; AI_REDACT( &nbsp; &nbsp; &nbsp; &nbsp; input &nbsp; &nbsp; &nbsp;=&gt; original_text, &nbsp; &nbsp; &nbsp; &nbsp; categories =&gt; ['NAME', 'NATIONAL_ID'] &nbsp; &nbsp; ) AS redacted_text FROM ( &nbsp; &nbsp; VALUES &nbsp; &nbsp; &nbsp; &nbsp; ('Customer John Smith applied for a loan. His SSN is 123-45-6789 and he lives in Boston, MA.'), &nbsp; &nbsp; &nbsp; &nbsp; ('Policy holder: Maria Gonzalez. National ID: 987-65-4321. Contact: maria.g@email.com'), &nbsp; &nbsp; &nbsp; &nbsp; ('Employee record: Name: David Lee, SSN: 456-78-9012, hired on 2021-03-15 in New York.') ) AS t(original_text)The categories argument lets us control which PII gets redacted. The output looks like this - note that the email in the second row was not redacted, because the EMAIL category was not included in the query arguments:REDACTED_TEXT"Customer [NAME] applied for a loan. His SSN is [NATIONAL_ID] and he lives in Boston, MA.""Policy holder: [NAME]. National ID: [NATIONAL_ID]. Contact: maria.g@email.com""Employee record: Name: [NAME], SSN: [NATIONAL_ID], hired on 2021-03-15 in New York."<figure id="figure-iwhnw13ldnmu2tkcn5"><img id="iwhnw13ldnmu2tkcn5" src="https://cdn.durable.co/blocks/11jY2Xy7IzEOkPQSQlwdg5m8XatuizA1xh0G1iGorG5whzRKNR2M7jPfy2JpAhQu.png" alt="" title="" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 3/2; object-position: 0% 100%;" data-corners="default" data-aspect="3:2" data-caption="Example of AI_TRANSLATE function" data-positionx="0%" data-positiony="100%" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/11jY2Xy7IzEOkPQSQlwdg5m8XatuizA1xh0G1iGorG5whzRKNR2M7jPfy2JpAhQu.png" data-mce-style="aspect-ratio: 3/2; object-position: 0% 100%;"><figcaption class="body-normal text-gray-500 mt-2">Example of AI_TRANSLATE function</figcaption></figure><h2>Aggregation text functions</h2> Both functions here perform the same action: summarizing text. Let's review the differences:<ul><li>AI_SUMMARIZE_AGG - takes one argument: the text column.</li><li>AI_AGG - takes two arguments: the text column and an instruction on how to summarize it.</li></ul>Which one you reach for depends on how much control you need over the output:SELECT &nbsp; &nbsp; AI_SUMMARIZE_AGG( &nbsp; &nbsp; &nbsp; &nbsp; original_text &nbsp; &nbsp; ) AS summary FROM ( &nbsp; &nbsp; VALUES &nbsp; &nbsp; &nbsp; &nbsp; ('awesome location, food is splendid'), &nbsp; &nbsp; &nbsp; &nbsp; ('loved the place'), &nbsp; &nbsp; &nbsp; &nbsp; ('nice cafe') ) AS t(original_text)SUMMARY This cafe is in an awesome location, serving splendid food, and is a lovely place to visit.SELECT &nbsp; &nbsp; AI_AGG( &nbsp; &nbsp; &nbsp; &nbsp; original_text, 'make a 3 words summary' &nbsp; &nbsp; ) AS summary FROM ( &nbsp; &nbsp; VALUES &nbsp; &nbsp; &nbsp; &nbsp; ('awesome location, food is splendid'), &nbsp; &nbsp; &nbsp; &nbsp; ('loved the place'), &nbsp; &nbsp; &nbsp; &nbsp; ('nice cafe') ) AS t(original_text)SUMMARY Great food location<figure id="figure-5bzpgmtht6mu2tnp6d"><img id="5bzpgmtht6mu2tnp6d" src="https://cdn.durable.co/blocks/3abrxRjUbGnmGrNtoJ8ricvzjLM3uppmSr13ha1GQaZol5s1IAt3s1IvX0S1YWJS.png" alt="" title="" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 16/9; object-position: 0% 50%;" data-corners="default" data-aspect="16:9" data-caption="Example of AI_AGG function" data-positionx="0%" data-positiony="50%" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/3abrxRjUbGnmGrNtoJ8ricvzjLM3uppmSr13ha1GQaZol5s1IAt3s1IvX0S1YWJS.png" data-mce-style="aspect-ratio: 16/9; object-position: 0% 50%;"><figcaption class="body-normal text-gray-500 mt-2">Example of AI_AGG function</figcaption></figure><h2>Functions that work with multiple formats</h2>Now we come to the large cluster of functions that work on text as well as other formats. Structuring them is no easy task. I propose grouping them into four clusters, ordered from the most specific task to the most generic:Derive a value from text or image<ul><li>AI_FILTER - answers TRUE or FALSE to a question, which you either embed in the text input or pass as a separate argument for images</li><li>AI_CLASSIFY - returns the user-defined category that best matches the text or image provided; it also supports a multi-category mode that returns a list </li></ul><figure id="figure-fjsygcbdgmu2tp8sv"><img id="fjsygcbdgmu2tp8sv" src="https://cdn.durable.co/blocks/5q4OP1fupUyi3edKMCy6pTfNQ2d1lfKGGo30760C6fRsja2MQp5778Hqo5U0DJC2.png" alt="" title="" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 16/9; object-position: 20% 50%;" data-corners="default" data-aspect="16:9" data-caption="Example of AI_CLASSIFY function" data-positionx="20%" data-positiony="50%" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/5q4OP1fupUyi3edKMCy6pTfNQ2d1lfKGGo30760C6fRsja2MQp5778Hqo5U0DJC2.png" data-mce-style="aspect-ratio: 16/9; object-position: 20% 50%;"><figcaption class="body-normal text-gray-500 mt-2">Example of AI_CLASSIFY function</figcaption></figure><h2>Vector embeddings from text, image, audio or video</h2><ul><li>AI_EMBED - creates a vector embedding that can be used for similarity comparison</li><li>AI_MULTI_EMBED - multimodal embeddings from text, image, audio or video; mainly used for semantic video search via the Marengo model</li><li>AI_SIMILARITY - calculates the similarity between two objects using the vector embeddings under the hood </li></ul><figure id="figure-shj7ec94j3mu2tqqbj"><img id="shj7ec94j3mu2tqqbj" src="https://cdn.durable.co/blocks/fVrFdtHfF8WzTbNFlq9geoAPjnZjIywgIyh83COufTLfOKueM7iGnnNQzwxBLAWP.png" alt="" title="" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="Example of AI_SIMILARITY function" data-positionx="undefined" data-positiony="undefined" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/fVrFdtHfF8WzTbNFlq9geoAPjnZjIywgIyh83COufTLfOKueM7iGnnNQzwxBLAWP.png" data-mce-style="aspect-ratio: 16/9; object-position: center center;"><figcaption class="body-normal text-gray-500 mt-2">Example of AI_SIMILARITY function</figcaption></figure>Extracting text from unstructured data<ul><li>AI_PARSE_DOCUMENT - extracts text and images from documents</li><li>AI_TRANSCRIBE - creates transcripts from audio and video files </li></ul>Answering a broad question<ul><li>AI_COMPLETE - generates a response to a question about text or image input (other formats in preview)</li><li>AI_EXTRACT - generates a response to a question about document, text or image input </li></ul><h2>What is the difference between AI_COMPLETE and AI_EXTRACT?</h2> As you can see in the last cluster, the two functions appear pretty close. Let’s review the main differences.Both functions support text and image inputs, but only AI_EXTRACT has generally available support for other document types (pptx, docx, pdf, etc.); in AI_COMPLETE, document, audio and video inputs are still in preview. AI_COMPLETE allows you to select the LLM model, while AI_EXTRACT does not let you pick a general-purpose model: its model argument only accepts a fine-tuned arctic-extract model.As a result, AI_COMPLETE allows you to modify the model’s parameters: temperature, top_p, max_tokens, guardrails. AI_EXTRACT can return a confidence score alongside the result. With AI_EXTRACT you specify the response format as a JSON-like schema, and the function always returns that shape. AI_COMPLETE returns free text by default, but it also accepts a response_format argument, taking either a JSON schema or a SQL type literal, and validates each generated token against it. With AI_COMPLETE, structure is opt-in, and it isn't available in every form: the PROMPT object variant, used for multi-file inputs, cannot take a schema. If you skip response_format and only describe the format in the prompt, the output is not guaranteed. As an example, let's use AI_EXTRACT. Here is our test prompt:SELECT AI_EXTRACT( &nbsp; &nbsp; $$I ordered the wireless headphones last month and overall I'm really satisfied with my purchase. The sound quality is excellent: crystal clear highs and solid bass. The noise cancellation feature works surprisingly well, even in a busy coffee shop. My only complaint is that it took about 3 weeks to arrive, which was longer than the estimated 5-7 business days. The packaging was great though, no damage at all. Battery life has been solid so far (about 8-9 hours with ANC on), and they're comfortable enough for extended use. The app integration is seamless. Would definitely recommend these, especially if you're on a budget and don't need the premium models.$$, &nbsp; &nbsp; {'issue_1': 'A first bullet point issue', 'issue_2': 'A second bullet point issue', 'issue_3': 'A third bullet point issue'} ) AS RESPONSE;The output follows the schema we defined:{ &nbsp; "error": null, &nbsp; "response": { &nbsp; &nbsp; "issue_1": "it took about 3 weeks to arrive", &nbsp; &nbsp; "issue_2": "The noise cancellation feature works surprisingly well, even in a busy coffee shop.", &nbsp; &nbsp; "issue_3": "None" &nbsp; } }<figure id="figure-qygqfgsa0imu2tvauk"><img id="qygqfgsa0imu2tvauk" src="https://cdn.durable.co/blocks/335CeJEUZ1DJ3vQaIYPIc07AGajzLpi3jzJUAkTEMczwFV2a5VHYNL6gRcjIO6sc.png" alt="" title="" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 16/9; object-position: undefined undefined;" data-corners="default" data-aspect="16:9" data-caption="Example of AI_EXTRACT function" data-positionx="undefined" data-positiony="undefined" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/335CeJEUZ1DJ3vQaIYPIc07AGajzLpi3jzJUAkTEMczwFV2a5VHYNL6gRcjIO6sc.png" data-mce-style="aspect-ratio: 16/9; object-position: undefined undefined;"><figcaption class="body-normal text-gray-500 mt-2">Example of AI_EXTRACT function</figcaption></figure><figure id="figure-yua95ciu5emu2tw495"><img id="yua95ciu5emu2tw495" src="https://cdn.durable.co/blocks/1b9EhgejSccme6GYkPK18odqTy86i9YPNTvp4npHW9K87xiBClir3ubELr8E2GBY.png" alt="" title="" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="Example of AI_COMPLETE function with structured output" data-positionx="undefined" data-positiony="undefined" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/1b9EhgejSccme6GYkPK18odqTy86i9YPNTvp4npHW9K87xiBClir3ubELr8E2GBY.png" data-mce-style="aspect-ratio: 16/9; object-position: center center;"><figcaption class="body-normal text-gray-500 mt-2">Example of AI_COMPLETE function with structured output</figcaption></figure><h2>AI Functions - accepted formats</h2>&nbsp;<figure id="figure-j0tw5qph1pmu2txww5"><img id="j0tw5qph1pmu2txww5" src="https://cdn.durable.co/blocks/38XviaIMv8aZj55Nq9r43nnZsivfEKP1nt6zzn4cJTNrOwgF9u6CyMvtl86nCpJQ.png" alt="" title="" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 16/9; object-position: 0% 10%;" data-corners="default" data-aspect="16:9" data-caption="" data-positionx="0%" data-positiony="10%" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/38XviaIMv8aZj55Nq9r43nnZsivfEKP1nt6zzn4cJTNrOwgF9u6CyMvtl86nCpJQ.png" data-mce-style="aspect-ratio: 16/9; object-position: 0% 10%;"></figure><h2>How much do Cortex AI functions cost?</h2>Pricing is based on AI Credit consumption, including both input and output tokens. Don't forget the cost of the warehouse running the SQL that calls the function. Snowflake recommends a warehouse no larger than MEDIUM for these calls: a bigger warehouse will not make the function faster, only the bill larger. Costs vary by function and by model. As an example, at the time of writing, the AI_COMPLETE function using Claude Sonnet 5 is billed at 1.20 AI Credits for input and 6.00 AI Credits for output per million tokens.To estimate costs for text-based inputs, use the helper function AI_COUNT_TOKENS, passing the function name and the input text as arguments.For the AI_EXTRACT function, tokens are counted per page of the document processed: each page counts as 970 input tokens, plus input/output tokens for the prompt, if the function uses one. AI_PARSE_DOCUMENT is billed per 1,000 pages processed rather than per token (0.68/3.66 AI Credits per 1,000 pages in OCR/LAYOUT mode).For transcription, each second of audio counts as 50 tokens, regardless of language or segmentation method. An hour of audio is therefore 180,000 tokens.<h2>Conclusion</h2>"Cortex AI functions" is an umbrella term covering a broad range of functionality. Navigating this landscape can feel daunting, but knowing the pros and cons of each function makes it much easier. If you want to know more about the rest of the Cortex family of features, see this <a data-mce-href="https://cloudhive.ai/blog/getting-started-with-snowflake-cortex" href="https://cloudhive.ai/blog/getting-started-with-snowflake-cortex">earlier article.</a>

New Navigation Guide: Snowflake Cortex AI Functions

Three HR consultants and startup founder reviewing candidate scorecards and hiring process flowcharts at conference table

Hey everyone, today I'm bringing you five non-obvious ways to use Snowflake CoCo. In the era of AI, we all use assistants like CoCo for various tasks, and I would like to inspire you to take some new approaches with it. Examples I bring will build on one another, but you can also use them independently. Let’s see what new things CoCo can do!<h2>Audit your account</h2>While we all spend a lot of time asking AI agents to do things for us, it is definitely a great idea to check their work, as well as ours (from those pre-AI days!). CoCo has a set of built-in skills to audit your account:<ul><li>Cortex Agents and Analyst: /semantic_studio</li><li>Governance and observability: /data-governance</li><li>Cost insights: /cost-intelligence</li><li>Data quality: /data-quality</li><li>AI preparedness: /ai-readiness-score</li><li>Your own skills: /skill-development … and the list is ever growing.</li></ul> I recommend using these skills as collections of best practices and checking your account against them. Chances are (and it definitely happened to me more than once) CoCo itself will not always follow every best practice from its own list on the first pass. After you’ve created something with it, run another pass to check the outputs!<figure id="figure-12m0oo8unvmtre39k0"><img id="12m0oo8unvmtre39k0" src="https://cdn.durable.co/blocks/23m9sp5MKsFuH44tZF1Y0pZmP3FTZUYShDEwb9nwdZTMyEtlUoJ23rI1J7ZhVKxJ.png" alt="" title="" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="Report I asked CoCo to produce based on the Governance Maturity Score in /data-governance skill" data-positionx="null" data-positiony="null" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/23m9sp5MKsFuH44tZF1Y0pZmP3FTZUYShDEwb9nwdZTMyEtlUoJ23rI1J7ZhVKxJ.png" data-mce-style="aspect-ratio: 16/9; object-position: center center;"><figcaption class="body-normal text-gray-500 mt-2">Report I asked CoCo to produce based on the Governance Maturity Score in /data-governance skill</figcaption></figure><h2>Reverse-engineer and learn from CoCo skills</h2> More often than not, the results of such assessments will come back with a score lower than expected. Is it bad? It depends. Your setup might be genuinely strong but failing against a hard-coded, overly generalized rule. Or the audit might have uncovered a real weakness. The point is this: whether you want to validate what's behind these audits or just satisfy your curiosity, you can see the backbone of the skills by copying their contents into your workspace to explore at leisure.<figure id="figure-725p60dzdqmtre3uvn"><img id="725p60dzdqmtre3uvn" src="https://cdn.durable.co/blocks/25bHpqWlo6GRBvbetKhbHKutBPVj9I2WrYoQoiIeqoAlsMoXzCXKUycHtlrxdJW8.png" alt="" title="" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="Copying the governance-maturity-score.md from the /data-governance skill" data-positionx="null" data-positiony="null" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/25bHpqWlo6GRBvbetKhbHKutBPVj9I2WrYoQoiIeqoAlsMoXzCXKUycHtlrxdJW8.png" data-mce-style="aspect-ratio: 16/9; object-position: center center;"><figcaption class="body-normal text-gray-500 mt-2">Copying the governance-maturity-score.md from the /data-governance skill</figcaption></figure><h2>Multi-agent teams and subagents</h2> Looking closely at the above list of tasks and skills, you could already draw inspiration for the next tip: parallelize CoCo’s work. Indeed, this is an option often overlooked: you can run a team of agents in a sub-workflow within a CoCo session, where the subagents share context and feed their outputs to a parent agent. All you need to do is ask for a team of agents or swarm. Note that it is best used for complex multi-stage tasks. Keep in mind that in some cases CoCo might still exercise its judgment to run with one agent, if it believes it will be more efficient that way.<figure id="figure-4unog6bfvjmtre4i20"><img id="4unog6bfvjmtre4i20" src="https://cdn.durable.co/blocks/27fppRCIlpYSuodHLXOcfNq4oNOlR2bpaNGuPp1Bx0yz5LnU4N3CeofzYreiAwRJ.png" alt="" title="" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="" data-positionx="null" data-positiony="null" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/27fppRCIlpYSuodHLXOcfNq4oNOlR2bpaNGuPp1Bx0yz5LnU4N3CeofzYreiAwRJ.png" data-mce-style="aspect-ratio: 16/9; object-position: center center;"></figure><figure id="figure-48ayne5h6mtre4kl9"><img id="48ayne5h6mtre4kl9" src="https://cdn.durable.co/blocks/2aMPDJ0FWbl0ypEhYVq2EDje9Gkdw0K2V9bW9IpkTHHsTBis5bBy0Spc9RKkWLBH.png" alt="" title="" class="object-cover bg-gray-100 " style="aspect-ratio: 16/9; object-position: center center;" data-corners="no-rounded" data-aspect="16:9" data-caption="A team of CoCo subagents runs checks in parallel and compiles results in one report (top and bottom of the screenshot shown, middle cropped for length)" data-positionx="null" data-positiony="null" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/2aMPDJ0FWbl0ypEhYVq2EDje9Gkdw0K2V9bW9IpkTHHsTBis5bBy0Spc9RKkWLBH.png" data-mce-style="aspect-ratio: 16/9; object-position: center center;"><figcaption class="body-normal text-gray-500 mt-2">A team of CoCo subagents runs checks in parallel and compiles results in one report (top and bottom of the screenshot shown, middle cropped for length)</figcaption></figure><h2>Synergy with Claude Code</h2> CoCo is great inside Snowflake, but how do you bring it to the broader world? Here is my favorite, and a bit of a hidden gem: <a data-mce-href="https://claude.com/plugins/snowflake-cortex-code" href="https://claude.com/plugins/snowflake-cortex-code">Claude Code plugin</a>. Once the plugin is installed, you can let Claude and CoCo discuss the approach and execution without breaking your flow or switching between interfaces. Why? - <a data-mce-href="https://www.snowflake.com/en/blog/engineering/data-eng-bench-data-engineering-agent-benchmark/" href="https://www.snowflake.com/en/blog/engineering/data-eng-bench-data-engineering-agent-benchmark/" target="_blank" rel="noopener">Let the new data-eng-bench</a> benchmark answer this question.<figure id="figure-n1j0rbw1rhmtre5uj7"><img id="n1j0rbw1rhmtre5uj7" src="https://cdn.durable.co/blocks/2dvkfBBbH80LMntj8smhRvCpP18o8lx9IQ41K7Sd6hBO5aXdnpmnh9Y6u18RZvgv.png" alt="" title="" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 1/1; object-position: center center;" data-corners="default" data-aspect="1:1" data-caption="Working in Claude to make a weekly report out of previously explored audits with skills" data-positionx="null" data-positiony="null" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/2dvkfBBbH80LMntj8smhRvCpP18o8lx9IQ41K7Sd6hBO5aXdnpmnh9Y6u18RZvgv.png" data-mce-style="aspect-ratio: 1/1; object-position: center center;"><figcaption class="body-normal text-gray-500 mt-2">Working in Claude to make a weekly report out of previously explored audits with skills</figcaption></figure><h2>Using CoCo in other apps</h2> As a cherry on top, the last surprising use of CoCo is in your own apps! Cortex <a data-mce-href="https://docs.snowflake.com/en/user-guide/cortex-code-agent-sdk/cortex-code-agent-sdk" href="https://docs.snowflake.com/en/user-guide/cortex-code-agent-sdk/cortex-code-agent-sdk" target="_blank" rel="noopener">Code Agent SDK (in Preview)</a> is meant exactly for this: incorporating CoCo agentic logic into your Python/TypeScript applications. As simple as pip install cortex-code-agent-sdk. Below is an example of a script running three CoCo agents from the earlier example to compile an audit report.""" Snowflake Account Health Swarm (Demo) ===================================== """import asyncio import json from cortex_code_agent_sdk import query, AssistantMessage, CortexCodeAgentOptionsJSON_OUTPUT_RULES = ( &nbsp; &nbsp; "Respond with ONLY a JSON object: no prose, no markdown, no code fences, " &nbsp; &nbsp; "nothing before or after it. Schema: " &nbsp; &nbsp; '{"score": &lt;1-10 integer&gt;, "findings": [&lt;short string&gt;, ...], ' &nbsp; &nbsp; '"recommendations": [&lt;short string&gt;, ...]}. ' &nbsp; &nbsp; "findings and recommendations must be flat arrays of short strings, not nested objects." )AGENTS = { &nbsp; &nbsp; "governance": f"Audit governance: check masking policies, row-access policies, tag coverage, and privileged roles using SNOWFLAKE.ACCOUNT_USAGE. {JSON_OUTPUT_RULES}", &nbsp; &nbsp; "cost": f"Analyze costs: check credit usage (30d/90d), top warehouses, weekly trend, and storage using SNOWFLAKE.ACCOUNT_USAGE. {JSON_OUTPUT_RULES}", &nbsp; &nbsp; "quality": f"Assess data quality: check table freshness, empty tables, query failure rates, and recurring errors using SNOWFLAKE.ACCOUNT_USAGE. {JSON_OUTPUT_RULES}", } def extract_json(text: str) -&gt; dict | None: &nbsp; &nbsp; """Pull the first {...} JSON object out of free-form text, if any.""" &nbsp; &nbsp; start, end = text.find("{"), text.rfind("}") &nbsp; &nbsp; if start == -1 or end == -1: &nbsp; &nbsp; &nbsp; &nbsp; return None &nbsp; &nbsp; try: &nbsp; &nbsp; &nbsp; &nbsp; return json.loads(text[start:end + 1]) &nbsp; &nbsp; except json.JSONDecodeError: &nbsp; &nbsp; &nbsp; &nbsp; return None async def run_agent(name: str, prompt: str) -&gt; dict: &nbsp; &nbsp; """Send a prompt to one Cortex Code agent and return its parsed JSON result.""" &nbsp; &nbsp; output = [] &nbsp; &nbsp; async for msg in query( &nbsp; &nbsp; &nbsp; &nbsp; prompt=prompt, &nbsp; &nbsp; &nbsp; &nbsp; options=CortexCodeAgentOptions( &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; allowed_tools=["SQL"], &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; # Both flags are required together, by design, as a two-key safety &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; # gate: permission_mode="bypassPermissions" states the *intent* to &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; # skip permission checks, but is a no-op on its own. &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; # allow_dangerously_skip_permissions=True is the separate &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; # confirmation that actually activates the bypass. This prevents &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; # permissions from being skipped by accident via a single flag. &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; permission_mode="bypassPermissions", &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; allow_dangerously_skip_permissions=True, &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; # Cost/quality agents sometimes burn a turn recovering from a &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; # blocked tool call before they even start querying; 10 wasn't &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; # always enough room left to also comply with the JSON-only &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; # output format at the end. &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; max_turns=15, &nbsp; &nbsp; &nbsp; &nbsp; ), &nbsp; &nbsp; ): &nbsp; &nbsp; &nbsp; &nbsp; if isinstance(msg, AssistantMessage): &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; output.extend(block.text for block in msg.content if hasattr(block, "text"))&nbsp; &nbsp; text = "".join(output) &nbsp; &nbsp; return extract_json(text) or {"agent": name, "raw": text} async def main(): &nbsp; &nbsp; print("Launching swarm: governance, cost, quality...\n")&nbsp; &nbsp; results = await asyncio.gather( &nbsp; &nbsp; &nbsp; &nbsp; *(run_agent(name, prompt) for name, prompt in AGENTS.items()) &nbsp; &nbsp; )&nbsp; &nbsp; max_console_findings = 3 &nbsp; &nbsp; for name, result in zip(AGENTS, results): &nbsp; &nbsp; &nbsp; &nbsp; findings = result.get("findings", []) &nbsp; &nbsp; &nbsp; &nbsp; print(f"[{name.upper()}] Score: {result.get('score', '?')}/10") &nbsp; &nbsp; &nbsp; &nbsp; for finding in findings[:max_console_findings]: &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; print(f" &nbsp;- {finding}") &nbsp; &nbsp; &nbsp; &nbsp; remaining = len(findings) - max_console_findings &nbsp; &nbsp; &nbsp; &nbsp; if remaining &gt; 0: &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; print(f" &nbsp;... +{remaining} more in swarm_report.json") &nbsp; &nbsp; &nbsp; &nbsp; print()&nbsp; &nbsp; with open("swarm_report.json", "w") as f: &nbsp; &nbsp; &nbsp; &nbsp; json.dump(results, f, indent=2) &nbsp; &nbsp; print("Full report: swarm_report.json") if __name__ == "__main__": &nbsp; &nbsp; asyncio.run(main())<figure id="figure-d9g54082oqmtre8jc9"><img id="d9g54082oqmtre8jc9" src="https://cdn.durable.co/blocks/317buc2gpD1OxhE0EP0zu10bTYOadTJOpkGUz7FAEBjntNO7qdartT1ZL7Mk0Sgw.png" alt="" title="" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="Results of the audit script on the author’s demo account" data-positionX="null" data-positionY="null" data-fullWidth="false"/><figcaption class="body-normal text-gray-500 mt-2">Results of the audit script on the author’s demo account</figcaption></figure><h2>Conclusions</h2>I hope this article sparked new ideas and pushed the limits of what you do with CoCo. If you'd like a refresher on CoCo billing, check out the <a data-mce-href="https://cloudhive.ai/blog/getting-started-with-snowflake-cortex" href="https://cloudhive.ai/blog/getting-started-with-snowflake-cortex" target="_blank" rel="noopener">previous article,</a> which covers all Cortex features. And remember: CoCo's capabilities expand every month, so there's more to come!

Five Ways to Use Snowflake CoCo You Did Not Consider

Diverse recruitment team analyzing hiring data and market intelligence reports during collaborative meeting in modern conference room

In this article, we are going to explore Snowflake Cortex Search and two of its applications that open up some unique options in the enterprise context. We will dive into the details of how to use it, along with some use cases I hope will spark ideas. But first, let's cover the basics.<h2>How does Cortex Search work?</h2>Snowflake Cortex Search is a fully managed hybrid search service. It combines keyword search and semantic search using vector embeddings over text columns in your tables. Cortex Search indexes those text columns and creates vector embeddings for them. This lets you search data for semantic similarity in addition to classic keyword matching. When you query it, Cortex Search returns row-level data containing the searchable columns and any other columns you specify, ranked by the relevance of the indexed columns to the query.<h2>How to create Cortex Search?</h2> You can create a Cortex Search service in Snowsight or using SQL. Important elements to specify include:<ul><li>indexed columns - these are the ones the search runs on</li><li>attributes - these are the columns you can filter results on at query time</li><li>target lag - this controls the freshness of the index, and therefore of the results </li></ul>Let's look at an example: the CUSTOMER_CASES table, which contains the TOPIC, REQUEST, and AGENT_RESPONSE columns describing specific issues. We would like to index TOPIC and REQUEST, make CASE_ID filterable as an attribute, and return CASE_ID and AGENT_RESPONSE alongside the search results.<figure id="figure-t4wmnsc4zmtisltzo"><img id="t4wmnsc4zmtisltzo" src="https://cdn.durable.co/blocks/207U5z7ssr1df58vRnNHR7k7nd4nwAcdi9rVf9rvOOaIPptRED61dekoFh3tT0RV.png" alt="" title="" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="" data-positionx="undefined" data-positiony="undefined" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/207U5z7ssr1df58vRnNHR7k7nd4nwAcdi9rVf9rvOOaIPptRED61dekoFh3tT0RV.png" data-mce-style="aspect-ratio: 16/9; object-position: center center;"></figure>Adding columns for text and vector indexing in the Snowsight UI<figure id="figure-ppfamplj4mtisory6"><img id="ppfamplj4mtisory6" src="https://cdn.durable.co/blocks/2aAgmAPppAZ47qggwdzLGY5dsJSRH7NFcOYwJyYochysMyJiWS4Sl0RWJSlQxy6b.png" alt="" title="" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="" data-positionx="undefined" data-positiony="undefined" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/2aAgmAPppAZ47qggwdzLGY5dsJSRH7NFcOYwJyYochysMyJiWS4Sl0RWJSlQxy6b.png" data-mce-style="aspect-ratio: 16/9; object-position: center center;"></figure>Adding the columns returned by the service in the Snowsight UI<figure id="figure-woyo7mr56qmtisp5zn"><img id="woyo7mr56qmtisp5zn" src="https://cdn.durable.co/blocks/23toUUhyKYulle36badIfpeTsQKMiwiIfJvcgMZmgEnWS3s7e9jzbW3N6qBgFHfK.png" alt="" title="" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="" data-positionx="undefined" data-positiony="undefined" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/23toUUhyKYulle36badIfpeTsQKMiwiIfJvcgMZmgEnWS3s7e9jzbW3N6qBgFHfK.png" data-mce-style="aspect-ratio: 16/9; object-position: center center;"></figure>Adding target lag in the Snowsight UIIf you prefer the SQL path, here is the complete statement to achieve the same result.CREATE OR REPLACE CORTEX SEARCH SERVICE DEMO_DB.PUBLIC.CUSTOMER_CASES_SEARCH &nbsp; TEXT INDEXES TOPIC, REQUEST &nbsp; VECTOR INDEXES TOPIC (model='snowflake-arctic-embed-m-v1.5'), REQUEST (model='snowflake-arctic-embed-m-v1.5') &nbsp; ATTRIBUTES CASE_ID &nbsp; WAREHOUSE = COMPUTE_WH &nbsp; TARGET_LAG = '1 hour' AS ( &nbsp; SELECT &nbsp; &nbsp; CASE_ID, &nbsp; &nbsp; TOPIC, &nbsp; &nbsp; REQUEST, &nbsp; &nbsp; AGENT_RESPONSE &nbsp; FROM DEMO_DB.PUBLIC.CUSTOMER_CASES );With the service created, we can look at the two main ways to use Cortex Search in the enterprise context.<h2>First use case: advanced search in your data</h2> The benefit of hybrid search is the ability to find results without knowing the exact wording. Taking our newly built search as an example, let’s say we would like the Customer Excellence team to be able to search existing cases while troubleshooting the customer issues they are working on. Existing cases often capture vague customer requests: the customer does not know the root cause or the exact product name, so they start by describing the issue and its symptoms. At this point, we would like a Customer Excellence representative to have a search bar where they can enter those symptoms and surface other cases with a similar issue, along with how other agents responded to them:SELECT PARSE_JSON( SNOWFLAKE.CORTEX.SEARCH_PREVIEW( 'DEMO_DB.PUBLIC.CUSTOMER_CASES_SEARCH', '{ "query": "charged twice", "columns": ["CASE_ID", "TOPIC", "REQUEST", "AGENT_RESPONSE"], "limit": 10 }' ) )['results'] AS results;As you would expect, the results are ranked by relevance, so this is a good point to look at how Cortex Search scoring works.<h2>How to improve Cortex Search scoring?</h2> The Cortex Search score combines vector similarity, text matching, and semantic reranking. By default, all three contribute equally, but you can configure their weights to tune the ranking. For example, if text matching matters much more in your use case, you can pass a scoring_config with your query like the following:"scoring_config": { &nbsp; "weights": { &nbsp; &nbsp; "texts": 3, &nbsp; &nbsp; "vectors": 1, &nbsp; &nbsp; "reranker": 1 &nbsp; } }If you use the same configuration repeatedly, you can save it as a named scoring profile with ALTER CORTEX SEARCH SERVICE … ADD SCORING PROFILE and reference it with the scoring_profile parameter instead.Similarly, you can boost specific columns in a multi-index search, on the text index, the vector index, or both. For example, if you want matches to come mainly from one column and use a second one as a fallback, you can set up the following configuration:"scoring_config": { &nbsp; "functions": { &nbsp; &nbsp; "text_boosts": [ &nbsp; &nbsp; &nbsp; {"column": "TOPIC", "weight": 1}, &nbsp; &nbsp; &nbsp; {"column": "REQUEST", "weight": 3} &nbsp; &nbsp; ], &nbsp; &nbsp; "vector_boosts": [ &nbsp; &nbsp; &nbsp; {"column": "TOPIC", "weight": 1}, &nbsp; &nbsp; &nbsp; {"column": "REQUEST", "weight": 3} &nbsp; &nbsp; ] &nbsp; } }Lastly, if your data includes numeric or time-based relevance signals, such as rankings or recency timestamps, you can factor those in as well, using numeric_boosts and time_decays in the scoring_config.Now, with enterprise search sorted, it is time to give this use case a new level of intelligence: plug the results of Cortex Search into AI.<h2>Second use case: use Cortex Search as a tool in Cortex Agent</h2> Imagine the same department, Customer Excellence, working through customer cases. But now, instead of colleagues pulling the example cases and going through them one by one in search of a possible solution, we delegate this preparation to an AI agent. This is where Cortex Agent comes into play. You can plug Cortex Search into it as a tool, enabling both search and analysis of the results. On top of that, you can run analytics across your whole corpus of unstructured data, asking questions like “what were the most frequent case topics in the last two months?” This capability has recently been rolled out as a Preview feature named <a data-mce-href="https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-agents-analytical-search" href="https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-agents-analytical-search" target="_blank" rel="noopener">Analytical search.</a><figure id="figure-zpxi16u19mtisppde"><img id="zpxi16u19mtisppde" src="https://cdn.durable.co/blocks/27CYJPBH9Nhcfqp4MSrWmUiAmkJnzOT76cSavwOhLBsTQHlFj0ysSwXMVeQ6VuHa.png" alt="" title="" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 1/1; object-position: center center;" data-corners="default" data-aspect="1:1" data-caption="" data-positionX="undefined" data-positionY="undefined" data-fullWidth="false"/></figure>You can add Cortex Search as a Cortex Agent tool in the Snowsight UIImportantly, you should always include a relevant tool description so the agent uses Cortex Search correctly, and spell out when the tool should be used in the orchestration instructions. You can read more about best practices for setting up Cortex Agents in my previous article. Here is an example of a Cortex Search tool description matching the above use case:Use this tool to search through historical customer support cases to find similar past issues and their resolutions. This tool supports two modes:1. **Standard retrieval**: Find specific past cases that are similar to the current customer issue, including the resolution steps taken. 2. **Analytical search**: Answer aggregate questions about the case corpus, such as identifying the most frequent case topics over a time period, common resolution patterns, or trending issues.Use standard retrieval when the user describes a specific customer problem and needs example resolutions. Use analytical search when the user asks about trends, frequencies, distributions, or patterns across many cases.<h2>How much does Cortex Search cost?</h2> Cortex Search pricing has a few components:Serving compute: a Snowflake-managed resource that incurs a fixed cost while the service is available to respond to queries. At the time of writing, and for all figures below, that is 6.3 AI Credits per GB of indexed data per month. The warehouse that runs during indexing, query orchestration, and refreshes incurs Platform Credits per hour, for example 1 Platform Credit per hour for an XS warehouse. AI_EMBED token usage for creating the initial embeddings and incremental updates of the indexed text varies in price depending on the embedding model. For example, using snowflake-arctic-embed-l-v2.0 costs 0.05 AI Credits per million tokens. Storage for the materialized index is a variable cost per TB per month based on region, for example $23.00 per TB per month in AWS US East. Note that batch search incurs different serving costs and an extra cost for query embedding. Note also that enabling Analytical search on a Cortex Search tool inside a Cortex Agent adds AI Function costs on top.<h2>Conclusions</h2> Snowflake Cortex Search works on its own as an enterprise search engine, and it also powers agentic search and analytics. It has multiple options under the hood for fine-tuning scoring quality. Given the cost, I recommend starting your testing on a smaller corpus of text. I hope these applications inspire you to keep experimenting and building with Snowflake!

Cortex Search – Two Applications for Enterprise Use

HR consultant reviewing recruitment metrics and candidate profiles on documents at minimalist desk

Semantic View Creation Guide: Autopilot or CoCo, or both?

Diverse healthcare professionals and recruiters collaborating in modern glass office during career placement consultation meeting

<h2>Intro</h2>CloudHive is bringing together experts from across the Snowflake community to share hands-on experience, technical insights and different perspectives from the field.In the latest article in our practitioner series, we take a closer look at Cortex Agent in practice - where it works well, where it can fall short and what teams should understand before building it into real-world workflows.<a data-mce-href="https://cloudhive.ai/blog/snowflake-cowork-3-new-features-every-analyst-should-know" href="https://cloudhive.ai/blog/snowflake-cowork-3-new-features-every-analyst-should-know" target="_blank" rel="noopener" data-mce-selected="inline-boundary">In a previous article</a>, we looked at the wider Cortex portfolio, including Cortex Analyst and Cortex Search as two of the main building blocks behind Cortex Agent.The series brings together practitioners from across the Snowflake ecosystem to share practical perspectives on architecture, engineering and AI on the platform.<h2>What is Cortex Agent?</h2>Cortex Agent is a Snowflake-native platform for creating agents, integrated with its other components, so you don't have to worry about connecting the pieces together. Snowflake takes care of the infrastructure as well as the logic of the agent when it comes to the underlying orchestration loop, runtime, build, and operations. While Snowflake manages all of these, you remain in control of the agent settings such as orchestration instructions and response instructions, as well as tools, skills, model selection, and more. The goal of Cortex Agent is to give you a ready-to-use platform into which you can plug your agent logic and your structured and unstructured data, then work with the resulting agent as your analytical companion. I'm going to share five tips and tricks that worked for me when setting up Cortex Agents. Without further ado, let's go.<figure id="figure-h6u2ucw0fmsmxvric"><img id="h6u2ucw0fmsmxvric" src="https://cdn.durable.co/blocks/dUK2ptnzsSbbopkDFT1TEKFpGZIYNuCWC4tM1IZDVAqaN10tuxMtxU2y2dElr6lg.png" alt="Configuration page of Cortex Agent" title="Configuration page of Cortex Agent" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 1/1; object-position: center center;" data-corners="default" data-aspect="1:1" data-caption="Configuration page of Cortex Agent" data-positionx="undefined" data-positiony="undefined" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/dUK2ptnzsSbbopkDFT1TEKFpGZIYNuCWC4tM1IZDVAqaN10tuxMtxU2y2dElr6lg.png" data-mce-style="aspect-ratio: 1/1; object-position: center center;"><figcaption class="body-normal text-gray-500 mt-2">Configuration page of Cortex Agent</figcaption></figure><h2>How does Cortex Agent work?</h2>Before we begin, let’s look under the hood of the orchestration, because it will be important for the first few tips. It is important to differentiate this implicit orchestration layer from the explicit orchestration instructions (the similar naming can be confusing). Behind the scenes, a Cortex Agent runs on a loop with three steps.<ol><li>Planning: the agent interprets the question, splits it into parts, and chooses which tools to use.</li><li>Tool use: the agent applies the chosen tool to answer the question, or part of it. This can be Cortex Analyst, Cortex Search, a custom tool such as a UDF or stored procedure, a web search, an MCP server, or a skill.</li><li>Reflection: the agent reviews the results it got back and judges whether to retry or respond with what it has.</li></ol><h2>Tip 1: You can't control the loop directly, but you can rein it in</h2>This loop brings us to the first caveat and the first tip: the above three steps aren't something you can directly control. There's no setting in Snowflake that lets you, for example, set how many loops run within a single request.The best approach I've found is to limit the number of loops through the agent's orchestration instructions. The agent loop itself remains something of a black box, but this one plain-language instruction “do not use more than X loops to get the answer”, placed in the orchestration instructions, lets you cap how many tools the agent plans to use. If you want to reduce this further, the next tip will help even more.<h2>Tip 2: Limits you can actually set to control the agent</h2> Another way to control the agent's response is to set explicit limits. Within the Agent Budget Configuration setting, you can set two limits directly:● &nbsp; &nbsp; &nbsp;Time Limit (seconds)● &nbsp; &nbsp; &nbsp;Token Limit <figure id="figure-bl0pj0oyfdmsmxx6fy" class=""><img id="bl0pj0oyfdmsmxx6fy" src="https://cdn.durable.co/blocks/15mgb2gsFxmm0icBxyHOnOyWlkGrZ3DjorZG27t5izPQ4ypTxDJ6EWU5BAwfTR5z.png" alt="Default recommended settings for Time and Token Limits in Cortex Agent" title="Default recommended settings for Time and Token Limits in Cortex Agent" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="Default recommended settings for Time and Token Limits in Cortex Agent" data-positionx="null" data-positiony="null" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/15mgb2gsFxmm0icBxyHOnOyWlkGrZ3DjorZG27t5izPQ4ypTxDJ6EWU5BAwfTR5z.png" data-mce-style="aspect-ratio: 16/9; object-position: center center;"><figcaption class="body-normal text-gray-500 mt-2">Default recommended settings for Time and Token Limits in Cortex Agent</figcaption></figure>Both limits apply on a per-response basis. Snowflake recommends at least 5 minutes as a default for best results, but depending on the complexity of your agent, simpler workflows may work fine with a lower limit. In my tests, however, setting the time limit to the minimum value of 1 second does not cause the agent to generate a response immediately, but it does significantly speed things up. My assumption from observing the behavior change is that the agent has a minimum thinking path it goes through, which reduces to a relatively naive approach with the strict limit. But in simple workflows it can be rather useful, so I encourage you to experiment with it.Also worth noting: token limits are only applicable at the orchestration level and do not include tool tokens. Similarly, in my tests, setting a token limit to the minimum value makes the agent invoke all steps explicitly and output all results of its thought process instead of conducting internal orchestration. In both cases, setting the parameter to a higher value results in longer but deeper thinking.<h2>Tip 3: Tell the agent explicitly which tool to use for which task</h2>Probably the most universal tip in this article: what can you do to improve the accuracy of the agent's response? An AI is only as good as the data feeding it, of course, but improving the data itself is a separate conversation we won't get into here.What's worth continuing to iterate on are the instructions you give the agent, specifically the orchestration instructions, which control how the agent selects tools and what sequence of actions it takes. The goal of these instructions is to make sure the agent knows which tool is right for which task.Say you have an agent that handles both structured and unstructured data. You want Cortex Analyst selected for structured-data questions and Cortex Search selected for unstructured-data questions, based on the type of question asked. You can teach the agent to predict which tool to use by writing this explicitly into the orchestration instructions, so it doesn't spend effort guessing which tool is most useful each time.&lt;aside&gt; &nbsp;💡As a rule of thumb: questions with a qualitative component should be answered by Cortex Search (for example, "What is our return policy?" or "How do we handle returns?"). Questions with a quantitative element, such as "How many returns did we have last quarter?", are better suited to Cortex Analyst.&lt;/aside&gt;Don't be shy about making this explicit in your agent's instructions. Include example questions and word markers that identify tools (“Sales” → Analyst, “Policy” → Search, etc.). Routing questions to the right tool is the single most useful thing you can put in your orchestration instructions: it reduces the time the agent spends selecting a tool and increases the accuracy of its behavior.<figure id="figure-zn9aztlgepmsmxz5y9"><img id="zn9aztlgepmsmxz5y9" src="https://cdn.durable.co/blocks/13Z6xoFvld0Bnax9kMqsP9bR1vbqDTeqHFYgU61DpFWmwA1K6tZrkAWyKCLzwHcM.png" alt="Orchestration instructions snippet for tool selection" title="Orchestration instructions snippet for tool selection" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="Orchestration instructions snippet for tool selection" data-positionx="undefined" data-positiony="undefined" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/13Z6xoFvld0Bnax9kMqsP9bR1vbqDTeqHFYgU61DpFWmwA1K6tZrkAWyKCLzwHcM.png" data-mce-style="aspect-ratio: 16/9; object-position: center center;"><figcaption class="body-normal text-gray-500 mt-2">Orchestration instructions snippet for tool selection</figcaption></figure><h2>Tip 4: Cut to the point and kill the kindness</h2>It might sound brutal, but the best thing you can teach your agent to do is to say “No.” In the world of LLM sycophancy, it is not a given that an agent can admit its own limits. If there's one thing worth putting in your response instructions, it's this: let the agent be explicit, and push back against the model's default tendency toward sycophancy. That's the headline example. More broadly, one of the most important things is to set the scope of questions the agent can and cannot answer, as well as explicitly include the boundaries for topics and actions it can assist with. Snowflake's own recommendation is to prefix all boundaries with the word “never” to exclude probabilistic interpretation of them and enforce the deterministic approach. So, if your agent is set up to work on datasets and provide insights about the data, some of the response instructions can be:● &nbsp; &nbsp; &nbsp;Never offer to draft an email to the customer● &nbsp; &nbsp; &nbsp;Never forecast sales data● &nbsp; &nbsp; &nbsp;Never invent data, always use only existing data● &nbsp; &nbsp; &nbsp;Never predict deal outcomes based on the dataAs a second part of this tip, I would recommend setting up a template for both the response format and contents, which would ensure repeatable behavior from the agent and reduce non-deterministic output. It is a general best practice to lead with a direct, short answer, and later expand with the details. In terms of formatting, I recommend you stick to your company's reporting guidelines. Don't forget citations: ask the agent to include references to all data, as this greatly increases trust (and reduces the risk of it inventing answers). One last word: since we all know LLMs are big fans of emojis, if you plan to have any, I recommend including them explicitly in the instructions, as this creates a recognizable pattern with repeated use.<figure id="figure-ulc6dtqzsmsmxzr1d"><img id="ulc6dtqzsmsmxzr1d" src="https://cdn.durable.co/blocks/11yUSpLBFJrXJEmohJMb2rI2u54BQJ1pTY7LOyIct9ibWjrGlM4k1NISpLmjLO64.png" alt="My demo Sales Agent getting emoji in the response instructions" title="My demo Sales Agent getting emoji in the response instructions" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="My demo Sales Agent getting emoji in the response instructions" data-positionx="undefined" data-positiony="undefined" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/11yUSpLBFJrXJEmohJMb2rI2u54BQJ1pTY7LOyIct9ibWjrGlM4k1NISpLmjLO64.png" data-mce-style="aspect-ratio: 16/9; object-position: center center;"><figcaption class="body-normal text-gray-500 mt-2">My demo Sales Agent getting emoji in the response instructions</figcaption></figure><h2>Tip 5: Determinism vs. Highest Quality Outcome</h2>This one was not obvious to me: when I started building Cortex Agents, it was generally recommended to pin the orchestration model of the agent to increase reproducibility. Yet, for some time now, Snowflake's official documentation has recommended selecting the “auto” model in production agents to achieve the best quality of answers. Some of the other Snowflake resources reference pinning the model for agents’ evaluations, which, if you think about it, makes a lot of sense. A word of caution: if the pinned model becomes unavailable (deprecated in Snowflake), Cortex Agent will start failing. And of course, if your model is set to “auto,” the cost of tokens will be subject to Snowflake's selection.<figure id="figure-bndt01of3msmy275o"><img id="bndt01of3msmy275o" src="https://cdn.durable.co/blocks/fBLGBGhTgCKblQEkjZL7taq8yjkztS80uNP5dbBOoP826URewNMdjWnkvjNZKKgu.png" alt="Model selection dropdown in the Cortex Agent Configuration setting" title="Model selection dropdown in the Cortex Agent Configuration setting" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 16/9; object-position: center center;" data-corners="default" data-aspect="16:9" data-caption="Model selection dropdown in the Cortex Agent Configuration setting" data-positionx="undefined" data-positiony="undefined" data-fullwidth="false" data-mce-src="https://cdn.durable.co/blocks/fBLGBGhTgCKblQEkjZL7taq8yjkztS80uNP5dbBOoP826URewNMdjWnkvjNZKKgu.png" data-mce-style="aspect-ratio: 16/9; object-position: center center;"><figcaption class="body-normal text-gray-500 mt-2">Model selection dropdown in the Cortex Agent Configuration setting</figcaption></figure><h2>Bonus: How to monitor the costs of Cortex Agent?</h2>Cortex Agent incurs costs on multiple levels: orchestration of the agent itself is charged per million tokens according to the consumption table. But bear in mind that each tool of the agent will incur its own costs. I prepared an example cost breakdown in an earlier article. Now let's look at cost control. Since Cortex Agent is a shared resource, it can be monitored using a shared resource budget:CALL AGENT_DEMO_DB.DEMO.AGENT_BUDGET!ADD_SHARED_RESOURCE('CORTEX AGENT',(SELECT SYSTEM$REFERENCE('CORTEX AGENT', 'AGENT_DEMO_DB.DEMO.SALES_ASSISTANT')));The mechanics of the shared resource budget have been reviewed in detail in my previous article about Snowflake CoWork.<h2>Conclusions</h2>Cortex Agents offer a flexible approach to setting up and maintaining serverless AI services in Snowflake. With default RBAC and Snowflake-native access to all the data and other Cortex services, it is the most intuitive way to bring your data under the AI umbrella. I hope these tips and tricks help you turn your agent from just good to great, and reduce the iterations you'll spend on settings, since I've already gone through them for you.

Make No Mistake: What Works and What Does Not in Cortex Agent

Three diverse professionals reviewing job postings on tablet during collaborative hiring discussion

CloudHive is bringing together experts from across the Snowflake community to share hands-on experience, technical insights and different perspectives from the field. In this latest article in the series, Dan Greenberg, Snowflake Data Platform Owner at Verisk, looks at an important change in Snowflake’s 2026_06 Behavior Change Bundle and what it means for organizations using transient tables as part of their backup strategy.Snowflake’s upcoming 2026_06 Behavior Change Bundle includes a feature change that every organization using Snowflake Backups should be aware of:&nbsp;Transient tables including transient dynamic tables will no longer be supported by Snowflake Backups.&nbsp;I understand Snowflake’s reasoning. Transient tables are intended for transitory or reproducible data that does not require the same recovery protections as permanent tables. However, this change could have serious consequences for organizations that currently include transient tables in long-term or immutable backup strategies.&nbsp;The goal of this article is not to criticize the change. It is to make sure Snowflake customers are aware of it, understand the impact, and are prepared before it takes effect.&nbsp;Before the change, transient tables could be backed up individually or included in a schema- or database-level backup. After the change:&nbsp;<ul><li>Creating a backup set directly against a transient table will fail with an error.</li><li>&nbsp;Existing table-level backup sets targeting transient tables will fail when attempting to add another backup.&nbsp;</li><li>Most concerning, a schema &nbsp;or database level backup will still succeed but Snowflake will silently exclude all transient tables.&nbsp;</li></ul>Those excluded tables will not be available during a restore because they were never captured. Existing backups that already contain transient tables are unaffected and can still be restored.&nbsp;<h2>What is a transient table?&nbsp;</h2> A transient table behaves much like a permanent Snowflake table: it persists across sessions, remains available until explicitly dropped, and can be accessed by other authorized users.&nbsp;The key difference is data protection.&nbsp;Transient tables:&nbsp;<ul><li>Have a maximum Time Travel retention period of one day.&nbsp;</li><li>Do not receive Snowflake’s seven-day Fail-safe protection.&nbsp;</li><li>Typically cost less to maintain because they do not incur Fail-safe storage costs.&nbsp;</li><li>Are best suited for staging data, intermediate transformations, ETL work tables, caches, and other data that can be independently reproduced.&nbsp;</li></ul>That last point is critical: transient tables should generally contain data that can be recreated without relying on Snowflake recovery. &nbsp;&nbsp;“Transient” and “must be backed up” are not mutually exclusive.&nbsp;It is tempting to assume that if a table was created as transient, its data cannot have a business, legal, or regulatory retention requirement.&nbsp;In the real world, that is not always true.&nbsp;A table may have originally been made transient to reduce Fail-safe storage costs. Years later, it may contain data needed for an audit, litigation, regulatory recordkeeping, historical analysis, or cyber-recovery. Business requirements change, applications evolve, and tables often live much longer than their original designers expected. The decision to avoid Fail-safe is not necessarily the same as the decision to exclude data from an immutable, long-term backup.&nbsp; Organizations should therefore rethink their use of transient tables. If there is any reasonable possibility that the data will need to be backed up, restored, audited, or retained for several years, a permanent table may be the safer design.&nbsp;The conversion is not painless&nbsp;Unfortunately, there is no simple command such as:&nbsp; ALTER TABLE my_table SET PERMANENT&nbsp;Snowflake does not allow the transient property to be changed after a table is created. Converting a transient table to a permanent table requires creating a new permanent table, copying the data, preserving grants and metadata, validating dependent objects, and then renaming or replacing the original table.&nbsp;At enterprise scale, this can become a significant migration effort.&nbsp;There is also a real cost tradeoff. Permanent tables receive a seven-day Fail-safe period, and the historical data maintained during that period incurs additional storage charges. Organizations that adopted transient tables to control storage costs must balance those new Fail-safe costs against their backup and retention requirements.&nbsp;<h2>Why this feature change deserves immediate attention&nbsp;</h2> Many large Snowflake environments have backup scopes containing tens of thousands of transient tables.&nbsp;Imagine discovering several years from now during an audit, investigation, legal request, or regulatory inquiry that required historical data is unavailable because those tables stopped being included in backups.&nbsp;<ul><li>The backup jobs continued to report success.&nbsp;</li><li>No database or schema level backup failed.&nbsp;</li><li>No warning clearly identified the missing tables.&nbsp;</li><li>The data simply was not captured.&nbsp;</li></ul>That would be a catastrophe and it is exactly the type of risk that can arise when a seemingly small feature change is missed.&nbsp;Make sure you are ready&nbsp;Before the 2026_06 bundle becomes enabled by default:&nbsp;<ol><li>Inventory all transient and transient dynamic tables within existing backup scopes.</li><li>Identify tables subject to business, legal, regulatory, audit, or cyber-recovery requirements.</li><li>Confirm that supposedly reproducible data can actually be regenerated within an acceptable timeframe.</li><li>Convert tables requiring long-term backups to permanent tables.</li><li>Estimate the resulting Fail-safe storage costs before migrating at scale.</li><li>Enable and test the behavior change bundle in a non-production account.</li><li>Add monitoring that compares expected backup objects with the objects actually captured.&nbsp; </li></ol>Again, I understand why Snowflake is making this change. Transient tables were designed for data that does not need permanent-table recovery protections.&nbsp;Just make sure your organization is aware of the change and ready for it.&nbsp; Table type is no longer only a storage-cost decision. It is also part of your backup, recovery, and compliance architecture.&nbsp;Snowflake documentation: <a data-mce-href="https://docs.snowflake.com/en/release-notes/bcr-bundles/2026_06/bcr-2360" href="https://docs.snowflake.com/en/release-notes/bcr-bundles/2026_06/bcr-2360" target="_blank" rel="noopener">2026_06 backup behavior change</a> | <a data-mce-href="https://docs.snowflake.com/en/user-guide/tables-temp-transient" href="https://docs.snowflake.com/en/user-guide/tables-temp-transient" target="_blank" rel="noopener">Transient tables</a> <a data-mce-href="https://docs.snowflake.com/en/user-guide/backups" href="https://docs.snowflake.com/en/user-guide/backups" target="_blank" rel="noopener">| Backups and immutable storage&nbsp;</a><h3>Talk to CloudHive&nbsp;</h3>Reviewing your Snowflake backup strategy or assessing how the 2026_06 feature change could affect your environment? CloudHive works with organisations across Snowflake architecture, data protection and platform governance to help identify risks and prepare for changes before they reach production.&nbsp;If you’d like to discuss your backup scopes, transient table strategy or a related topic you would like us to cover in a future article, <a data-mce-href="https://cloudhive.ai/contact" href="https://cloudhive.ai/contact" target="_blank" rel="noopener">get in touch.&nbsp;</a>

Snowflake’s 2026_06 Feature Change: Transient Tables Will Be Excluded from Backups

CloudHive

CloudHive is your North America and LATAM Snowflake-focused staffing partner.

We help accelerate Snowflake adoption and maximize the value of the Snowflake AI Data Cloud by delivering experienced, on-demand Snowflake contractors and specialized talent acquisition services.

Our exclusive focus on Snowflake gives us access to a deep network of certified contractors and professionals built through long-standing relationships, enabling us to deliver high-impact talent quickly and precisely.

Beyond contractors, we place experienced full-time hires across Data and AI Leadership, Technical Delivery, and Sales, ensuring organizations have the expertise to scale and succeed.

Switzer

Rufina

To offer you the best experience, this site uses cookies.

Blog

New Navigation Guide: Snowflake Cortex AI Functions

Five Ways to Use Snowflake CoCo You Did Not Consider

Cortex Search – Two Applications for Enterprise Use

Semantic View Creation Guide: Autopilot or CoCo, or both?

Make No Mistake: What Works and What Does Not in Cortex A...

Snowflake’s 2026_06 Feature Change: Transient Tables Will...