Local RAG Builder

Problem

Local RAG setups often stall on environment setup, model downloads, chunking quality, and retrieval noise: missing Python packages, failed embedding model fetches, split code blocks and tables, and multiple knowledge bases bleeding into each other. local-rag-builder turns these steps into a debuggable pipeline so an agent gets usable context before answering.

How It Works

• Setup: rag_env_setup.py checks Python 3.11+, installs chromadb, sentence-transformers, and langchain dependencies.
• Model and ingestion: downloads embedding models from multiple sources with retries and path correction; documents are chunked by text_splitter.py, stored in Chroma, and deduplicated by SM3 hash.
• Chunking and retrieval: fixed window, recursive, header hierarchy, sentence, and semantic chunking; GuardStack protects code, formulas, tables, and HTML; optional reranking and knowledge-base routing are available.
• Two modes: rag_skill.py retrieves only and lets the agent answer; rag_standalone.py retrieves, then calls LM Studio, Ollama, or vLLM to generate the answer.

Boundaries

Best for local private docs, Markdown/PDF material, and single-process knowledge bases. Native text formats work best; PDF/image OCR must be enabled; it does not directly support OpenAI/Cohere API embeddings; keep each base under about 50k entries and avoid multi-user concurrent writes.