Answers grounded in your own documents, with sources shown.

In the wild

Before

• —Generic chatbots either didn't know a business's own documents, or answered confidently with something made up when they weren't sure.
• —A fixed-size chunking approach split documents mid-sentence or mid-table, losing the structure that made a chunk meaningful on its own.
• —An answer with no source attached was a black box — no way to check it against the document it actually came from.

The Brain

Every document is split by heading, paragraph, list, table and code block — not a fixed character count — and converted into a 384-dimensional embedding stored in Qdrant with its source file, chunk number and heading path.

The Loop

A question is embedded and matched against the closest stored chunks; those chunks, not the raw question alone, are what a local Ollama/Gemma model answers from — with multi-turn conversation history carried across follow-up questions.

The Triage

What runs itself

A question matched confidently against the retrieved chunks gets a grounded answer immediately, with the source file and heading cited alongside it.

What reaches a person

Adding, removing or correcting a document in the library is a decision for whoever owns the knowledge base — the system re-embeds it, it doesn't decide what belongs there.

Where it stands

Delivered as an end-to-end RAG pipeline: questions are answered from the real document set with sources attached, new documents can be added without retraining anything, and follow-up questions keep their context.

Delivered with

• Semantic document chunking
• Vector search (Qdrant)
• Source-cited answers
• Multi-turn conversation history
• Local LLM (Ollama + Gemma)