October 31: Support for simulated tool calling, bug fixes

New Features

  • Graphlit now supports simulated tool calling for LLMs which don't natively support it, such as OpenAI o1-preview and o1-mini. Tool schema will be formatted into the LLM prompt context, and tool responses are parsed out of the JSON formatted response.

  • Given customer feedback, we have lowered the vector and hybrid thresholds used by the semantic search. Previously, some content at a low relevance was being excluded from the semantic search results. Now, more low-relevance content will be included in the results, used by the RAG pipeline. Reranking can be used to sort the search results for relevance.

Bugs Fixed

  • GPLA-3357: Not extracting all images from PDF, and should filter out single-color images.

Last updated