Skip to content
Building

AI · Retrieval-augmented generation

DocMind — a RAG system that answers from your documents

A general-purpose language model answers from memory. Ask one about your own reports and it sounds just as sure when it is guessing. DocMind works the other way round: it reads your documents first, finds the passages that matter, and only then writes an answer — with those sources attached.

Role
Solo build — data pipeline, retrieval, generation and interface
When
2026
Stack
Python · LLMs · Sentence embeddings · FAISS / Chroma · LangChain · Streamlit
System

Ingest

  1. DocumentsPDF · reports
  2. Chunkoverlapping
  3. Embedsentence vectors
  4. Vector indexFAISS / Chroma

Vector index → Retrieve · nearest passages by meaning

Ask

  1. Question
  2. Retrievetop-k passages
  3. Generategrounded LLM
  4. Answerwith citations

01

The problem

Teams keep what they know in PDFs, reports and spreadsheets that nobody has time to re-read.

A plain chatbot cannot see those files, and when it answers anyway it can invent facts that look real.

02

What I built

  • An ingestion pipeline that splits documents into overlapping chunks and turns each chunk into a sentence embedding.
  • A vector index for semantic search, so a question finds passages by meaning rather than by matching keywords.
  • A generation step that may only answer from the retrieved passages, and returns them as citations.
  • A chat interface with conversational memory, so follow-up questions keep their context.

03

How it works

  1. Step 01

    Ingest

    PDFs, reports and notes are parsed and split into overlapping chunks, so no sentence loses its context at a boundary.

  2. Step 02

    Embed and index

    Each chunk becomes a sentence embedding stored in a vector index (FAISS or Chroma).

  3. Step 03

    Retrieve

    A question is embedded the same way, and the passages closest to it in meaning are pulled back.

  4. Step 04

    Answer with sources

    The model writes the answer from those passages only, and every claim links back to the text it came from.

04

Outcome

  • Answers are grounded in your own text and show their sources, so anyone can check them in seconds.
  • In active development: the retrieval pipeline and the chat interface are the current focus.

What's next

  • An evaluation set of real questions with known answers, to measure accuracy before and after every change.
  • Hybrid search (keyword plus semantic), so exact figures like invoice numbers are found reliably.
  • A public demo.

Want this kind of work on your data?

Get in touch