DocuChat

• Splits your PDFs, embeds them with nomic-embed-text and stores them in PostgreSQL with pgvector.
• You ask in plain language and every answer cites its source documents.
• It runs fully locally with Ollama or with Gemini in the cloud, switched by an environment variable.

• Hybrid search. Vector search failed on short questions like “What is the PMBOK?”: the right passage did not even make the top-k. Combining it with PostgreSQL full-text search and fusing both with Reciprocal Rank Fusion moved that passage to first place.
• Tool-calling agent. For each question the model decides whether to search the documents or check the date. I compared Ollama and Gemini: both choose well; Ollama is free and private, Gemini polishes answers a little more.
• Observability. Every query stores provider, tools, retrieved passages and latency, to measure instead of guess.

recall@8: the right source shows up in the top 8 passages

60 %

questions in the evaluation set

15

model providers: local Ollama and Gemini

2

A first version of the evaluation gave 93%, but it counted generic passages that did not answer the question as hits. Once corrected, the real figure is 60%, with failures concentrated in broad conceptual questions.