Skip to content

rag-library.md

Latest commit

 

History

History
132 lines (87 loc) · 3.16 KB

rag-library.md

File metadata and controls

132 lines (87 loc) · 3.16 KB

High-Precision eBook RAG (SFTP Inbox)

Author: Mr. Watson 🦄 Date: 2026-02-08

Goal

Run a local high-precision RAG pipeline where documents are dropped via SFTP and indexed automatically.

Quick operations

⚠️ On-Demand Service: RAG library does NOT auto-start on boot. Use gpu-service to manage it:

gpu-service status                    # Check if running
gpu-service start rag                 # Start when needed (e.g., before uploading docs)
gpu-service stop rag                  # Stop when inbox is empty

See GPU Service Management for details.

Status check (when running):

systemctl status rag-library-ingest
journalctl -u rag-library-ingest -n 120 --no-pager
ls -lah /home/sftpuser/library_inbox /home/sftpuser/library_done /home/sftpuser/library_failed

Ingest folders (SFTP)

Upload new files to:

  • /home/sftpuser/library_inbox

Pipeline moves files to:

  • /home/sftpuser/library_done (indexed OK)
  • /home/sftpuser/library_failed (parse/index errors)

Supported formats

  • .pdf
  • .epub
  • .txt
  • .md

Pipeline service

Systemd service:

  • rag-library-ingest.service

Script:

  • /opt/rag-library/rag_pipeline.py

Virtual env:

  • /opt/rag-library/.venv

Data:

  • Chroma vectors: /opt/rag-library/data/chroma
  • File registry (SQLite): /opt/rag-library/data/registry.db

Precision setup (high)

  • Embeddings model: intfloat/multilingual-e5-base
  • Semantic chunking with overlap
  • Metadata per chunk (source hash, title, chunk index)
  • Cross-encoder reranker: cross-encoder/ms-marco-MiniLM-L-6-v2
  • Dedup by SHA-256
  • Incremental ingest (new/changed files only)

GPU (2026-02-18)

Upgraded eGPU to RTX 2070 Super 8GB (sm_75). All inference now runs on CUDA:

  • Embedding model (multilingual-e5-base): device='cuda' — ingest and query
  • Cross-encoder reranker (ms-marco-MiniLM-L-6-v2): device='cuda'

No CPU inference paths remain in rag_pipeline.py.

Environment config

/etc/rag-library.env

RAG_EMBED_MODEL=intfloat/multilingual-e5-base
RAG_RERANK_MODEL=cross-encoder/ms-marco-MiniLM-L-6-v2

Operations

# Service status
systemctl status rag-library-ingest

# Logs
journalctl -u rag-library-ingest -n 200 --no-pager

# Check indexed/failed counts
sqlite3 /opt/rag-library/data/registry.db "select status,count(*) from files group by status;"

# One-shot manual query
/opt/rag-library/.venv/bin/python /opt/rag-library/rag_pipeline.py query "tu pregunta"

Current status snapshot (2026-02-08)

  • service: active
  • indexed files: 2
  • failed files: 0
  • inbox watcher: active

(Use SQL command above for live numbers.)