Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.ai-stats.phaseo.app/llms.txt

Use this file to discover all available pages before exploring further.

Use this recipe when a web-search request is not behaving the way you expected and you want to see whether AI Stats accepted it, routed it correctly, and returned search results cleanly.

​
1. Make the search intent explicit

Do not rely on prompt wording alone. Include the search settings directly in the request so AI Stats knows you want web search. Check for:
  • native search tools in tools
  • web_search_options when the endpoint supports them

​
2. Confirm that a compatible provider was available

If the request is blocked or routed differently than you expected, inspect the request details for:
  • provider candidate details
  • provider enablement details
  • routing explanation
This tells you whether the provider list was narrowed because search support was missing.

​
3. Inspect usage counters

Successful search-backed requests should show search-related usage meters such as:
  • requested native web search tools
  • web search results
  • citations
These are the fastest signal that the request actually used search.

​
4. Inspect the search results section

For successful requests, open the request detail dialog and look for the Web search observability section. That section shows:
  • whether native search was used
  • normalized result count
  • normalized citation count
  • the extracted result records
  • the extracted citation records
Use this to compare provider behavior without reading raw payloads.

​
5. Keep these cases separate

Today, native search and provider capability gating are live. Search fallback planning exists separately. When debugging, keep these cases separate:
  1. native search request accepted and routed
  2. native search request accepted but no compatible provider remained
  3. native search request succeeded but returned weak or sparse citations
These are different problems and should not be debugged the same way.

​
6. What to compare when results look wrong

Compare across requests:
  • model ID
  • provider actually selected
  • request search options
  • result count
  • citation count
  • routing details
If result quality differs by provider, the normalized logs should still make that visible.
Last modified on May 19, 2026