RAGForge documentation
RAGForge is one toolkit for parsing, chunking, retrieval, grounded answers, evaluation, quantization, migration, and multi-agent coordination — exposed as an HTTP/JSON API so any language can use it. The core install has zero required dependencies.
Source of truth is the repo: github.com/samsuljahith/RagForge. Every page here links back to the code it describes.
Pipeline at a glance
Parse
Chunk
Retrieve
Answer
Start here
Quickstart
Install, build a knowledge base, run your first query in 5 minutes.
Architecture
How the modules connect via the plugin registry. Ingest and query flows.
CLI reference
Every ragforge subcommand and flag, sourced from cli.py.
HTTP API
Every endpoint, request body and response shape.
Python modules
Public classes and functions across parsing, chunking, pipeline, etc.
Evaluation
Golden datasets, metrics, A/B compare.
Multi-agent
Blackboard-based coordination. How agents share state.
Migration
Shadow-index a new embedding model, validate, then cut over.
What's in core
- Core — Document / Chunk data models and the plugin registry.
- Parsing — txt, md, html, pdf, and optional Docling for complex layouts.
- Chunking — fixed sliding window, structure-aware, and Docling chunker.
- Pipeline — embed + store + hybrid search (dense + BM25 via RRF) + cross-encoder rerank.
- Generation — grounded answers with citations; refuses on insufficient evidence. OpenAI / Anthropic / Ollama.
- Evaluation — hit_rate, MRR, precision@k, recall@k, faithfulness, answer relevance.
- Quantization — compress embeddings; measure tradeoff on your own data.
- Migration — shadow-index a new model, validate, atomic cutover.
- Coordination — multi-agent blackboard so agents share state instead of messaging.
- Tracing & UI — every query traced to SQLite;
ragforge uifor a local dashboard.