I've been frustrated by the complexity of modern RAG stacks. To run a simple document search, you usually need Docker, Pinecone/Milvus, an Embedding Model, and heavy dependencies like LangChain or Torch.
I wanted an architecture that was truly portable.
Introducing RAGdb (v1.0.6)
It’s an embedded, multimodal knowledge graph that lives entirely inside a single SQLite file.
The Novelty:
Instead of heavy embeddings, it uses a Hybrid Search Engine (TF-IDF Vectorization + Exact Substring Boosting) written in pure NumPy. This allows it to run on edge devices, CI/CD pipelines, or inside strict corporate environments where you can't spin up servers.
Key Features:
Zero Heavy Dependencies: The core is <30MB.
Portable Container: The .ragdb file contains the vectors, the metadata, the extracted text, and the search index. You can email the database to a colleague.
SOTA OCR: Optional support for ONNX-based OCR if you need to index images.
Incremental Ingestion: It hashes files and only re-processes changed documents.
Installation:
pip install ragdb
Code & Architecture:
https://github.com/abkmystery/ragdb
I’m looking for feedback on the retrieval architecture. I believe this "Single-File" approach is the missing link for local-first AI.
AbkMystery•1h ago