Agentic Local RAG for Your Company Knowledge

INDEXING · runs in the background whenever documents change

1. Your data

   files · mail · ERP · DMS

   AfaRAG connects to the systems you already use and watches them for changes. Nothing is copied to the cloud.

2. Parse & OCR

   layout · tables · scans

   Every document is read the way a person reads it: headings, tables and page numbers are kept, and scanned pages go through local OCR.

3. Chunk & enrich

   sections · page stamps

   Text is split along its own structure. Each chunk carries its section path, page number and folder, so a bare table of numbers still knows what it belongs to.

4. Embed

   dense + sparse vectors

   A multilingual embedding model turns each chunk into a meaning vector, and a sparse keyword vector is stored beside it for exact terms and part numbers.

5. Extract entities

   companies · people · values

   Companies, people, projects, amounts and dates are pulled out, merged across spellings, and every relation keeps the file and page that prove it.

6. Hybrid index

   vectors + keywords

   One vector store holds dense and sparse vectors for millions of chunks, with filters for project, document type, date and access rights.

7. Knowledge graph

   entities & relations

   Who worked with whom, on which project, documented where. The graph answers relationship questions that text search alone cannot.

8. Fact store

   structured records

   Contacts, invoices, bids and test results are kept as structured records, so counts, sums and comparisons are calculated, not guessed.

1. Local model stack

   all on your GPU server

   Every model runs on your own hardware. Nothing calls an external API, and each model can be swapped when a better open model appears.

QUESTION → ANSWER · every query, in seconds

1. Question

   chat · voice · API

   Ask in your own words and language: typed, spoken, with an image attached, or from another system through the API.

2. Rewrite & route

   intent → tools

   A small local model strips the small talk, fixes names, turns the question into a search phrase and decides which tools the question needs.

3. Retrieve

   up to 7 lanes, fused

   Semantic search, keyword search, exact identifiers, file paths, graph and fact lookups run in parallel. Their results are fused into the 100 best candidates.

4. Rerank & gate

   100 → ~8 passages

   A cross-encoder reads the question and each passage together. Weak passages are dropped, and the survivors are expanded to their full table or section.

5. Generate

   local LLM · cites sources

   The local LLM reasons over the selected passages only and writes the answer in the language you asked in, citing every claim.

6. Verify

   numbers · names · files

   Every number, name and file name in the answer is checked against the sources. Anything that cannot be found there is flagged under the answer.

7. Answer + sources

   page-exact citations

   You get the answer with the documents and pages it came from, one click away from the original file.

↻ Agent loop: evidence too thin? Plan, search again, call tools