Skip to content

Search and clustering

Note

🚜 This page is still under construction! 🚜

  • If you don't precisely know what you're looking for, start with a general search (e.g., "cases of cheating" or "types of environment issues"). Then, based on initial results, refine your query.
  • If you do know what you're looking for, feel free to provide lots of detail in your query; we use reasoning models to determine whether an agent run matches your query.
  • Use appropriate metadata filters to narrow the scope of your search.

Linking to search results

Warning

We're still ironing out some weird behavior when the base filter changes. You may encounter slowness if you share a completed search result, make the base filter less restrictive, and then open the link.

Search results are persisted and shareable: you can click the share button on both overall results and individual transcript blocks to get shareable links.

Retrieving search results from the SDK

The Python SDK exposes search results via get_search_results_for_id, given a FrameGrid ID (fg_id) and Dimension ID (dim_id).

results = client.get_search_results_for_id(fg_id, dim_id)

If you aren't sure which fg_id and dim_id to use, you can call client.list_framegrids() to find the right FrameGrid, then use client.list_attribute_searches(fg_id) to find the right dim_id corresponding to your search.

# Find the right FrameGrid
framegrids = client.list_framegrids()
print(framegrids)
fg_id = framegrids[0]["id"]  # An arbitrary FrameGrid

# Find the right dim_id
searches = client.list_attribute_searches(fg_id)
print(searches)
dim_id = searches[0]["dim_id"]  # An arbitrary Dimension

You'll get back a list of Attribute objects from get_search_results_for_id, which might look like:

[
    {
        'id': '23daf040-4b78-45d9-a2be-448c0c265275',
        'datapoint_id': '263b8da8-3d30-4d79-9a69-466b695a65cb',
        'attribute': 'agent uses ls',
        'attribute_idx': 0,
        'value': 'In [B2], the agent uses the `ls` command to list the contents of the current directory. This command returns "link.txt", showing there\'s a single text file in the working directory.'
    },
    {
        'id': '9c015cd3-838c-458d-a550-46a7efb6082c',
        'datapoint_id': '56f2d9f1-be00-49dc-b0cb-a1beeccef9b8',
        'attribute': 'agent uses ls',
        'attribute_idx': None,
        'value': None
    },
    ...
]