docs/how_to/tool_artifacts/

[giscus[bot]]

docs/how_to/tool_artifacts/

This guide assumes familiarity with the following concepts:

https://python.langchain.com/docs/how_to/tool_artifacts/

[Giambapisasale]

How to handle artifacts in create_react_agent with tools?

[carlosmontoliu]

Also really interested

[romanh24]

What is the real-world use case for the artifact?

[Mathbkj]

I'd also love to know that

[dabdine]

I typically use artifacts as "evidence" for a human to review. For example:

• Returning a raw API response object alongside the text provided to the LLM for further turns.
• Providing a screenshot of a website I scraped content from (i.e., you don't know if the actual text you scraped is the same as what was rendered in the page, so you can provide this as evidence so that a human can validate the LLM's thinking).
• Providing tabular data from some document (as the data source) for rendering / display in the UI separate from the LLM summarization of the data.

[dabdine]

Other examples:

• Links to specific websites you obtained some information from (rather than relying on the LLM hallucinating the tokens back out).

See what Perplexity does as an example. Their responses include images and links from the sources they cite. Consider those as artifacts that were a function of whatever research it did along the way as evidence/proof of its answer.

[dulnoan]

Is there an equivalent guide for MCP Servers?

For example the AI response usually gives a summary of the data, but it would be great to be able to get the complete dataset from API response i.e AI is instructed to return the top 10 from a sample of 50 but it would be great if the 50 can be returned as an artifact as well to the user.