# FLAMEHAVEN FileSearch

A self-hosted retrieval-augmented-generation search engine that indexes 34 file formats with hybrid BM25 and semantic search, and can call Gemini, OpenAI, Claude or a local Ollama model to answer from the retrieved passages.

FLAMEHAVEN FileSearch is a self-hosted RAG search engine meant to keep documents off third-party services. It combines BM25 keyword search with semantic vector search in a hybrid ranking mode, supports typo correction, and parses 34 file formats including PDF, Word, Excel, PowerPoint, RTF, HTML, CSV, LaTeX and images.

A quality gate scores each hybrid search result's confidence and can fall back to a keyword search or flag a low-confidence result rather than return it silently. The retrieval pipeline uses structure-aware chunking, two-level indexing, and sliding-window context enrichment, and it has a dedicated mode for ingesting an Obsidian vault's Markdown notes, including frontmatter, aliases, tags and wikilinks. Framework integrations for LangChain, LlamaIndex and similar libraries are imported only when needed.

It runs as a Docker container, is configured to call one of four large-language-model backends, Gemini, OpenAI, Claude or Ollama, by an environment variable, and can also be installed as a Python SDK or called over a REST API. It is MIT licensed.

Website: https://flamehaven.space/work/flamehaven-filesearch/