Implement external LLM usage #62
Conversation
Replaced external API dependency with efficient local search system: - Added relevance scoring with configurable thresholds - Implemented token limit enforcement (15k tokens) - Added regex escaping for safe pattern matching - Filtered overly broad queries with 50% threshold - Removed dotenv dependency - Added documentation caching system
|
@BurgessTheGamer is attempting to deploy a commit to the MeshJS Team on Vercel. A member of the Team first needs to authorize it. |
|
I’m concerned this change downgrades the retrieval quality. Replacing vector db with keyword search removes semantic understanding, i.e the system matches based on keywords and not concepts. f.e. "mint token" vs "create asset" are semantically same, because they use different wording they’ll no longer retrieve the same results. also If you're using llms.txt file under the hood, doesn't it make more sense to index the llms.txt file directly into the code editors? then we don't need of this mcp. |
|
i understand what youre saying, yeah this makes sense. now i feel dumb XD it would be nice if there was an endpoint for external LLMs to tap in to the vector db directly, basically giving your claude code the same info as the website chat |
BurgessTG
changed the title
|
Something along the lines of /search-docs endpoint that preserves your semantic vector search but returns raw documentation instead of generating answers |
2 participants
Replaced external API dependency with efficient local search system: