We’ve recently swapped our Agentic Tool decision model from Kimi K2 to Gemini Flash Lite.
Basically, the model has access to 5 available tools
- librarySearch - internal rag
- academicSearch - exa search for papers
- internetSearch - exa for internet
- youtubSearch - exa for youtube
- stop - this is an actual tool that the LLM can decide to call to stop the response
We typically encounter error 500 in the following scenario.
1st run → model decides a tool to call → executes tool
2nd run → tool call and tool result are added to history → 500 error sporadically
Unrelated - We’ve also discovered separately that implicit caching is unreliable in these cases.