Nex Open Brain Rag

What it does

• Delivers the full stack: models.py (768-dim vector schema with thoughts, connections, syntheses, reminders, conversation history), file_ingestor.py, embedding_service.py, vector_search.py, config.py, plus docker-compose.yml (pgvector/pg16 bound to localhost), Dockerfile and .env.example.
• Ingests folders of markdown and text notes: chunks at 1500 chars with 200 overlap, embeds locally with Ollama (nomic-embed-text, no per-call embedding cost), and stores a file header on the first chunk for traceability.
• Auto-extracts metadata per thought: summary, topics, people, sentiment, with a no-API fallback so ingestion works before any chat key is set.
• Answers questions via cosine-distance semantic search and a chat fallback chain: local Ollama, then Qwen, then Claude.
• Exposes the whole thing over a Telegram bot, including capture-on-the-fly: one message becomes a tagged, searchable thought.

Why use this

Your notes become a knowledge base you query in natural language, on hardware you own, with embeddings that cost nothing per call. The chat chain degrades gracefully (local model first, paid APIs only as needed), Postgres binds to 127.0.0.1 only, and all secrets live in env vars. Sized to run on a Raspberry Pi 5 and built from production patterns at a Belgian AI agency.

See it in action

Prompt:

I have years of markdown notes. I want to ask questions over them, self-hosted on my Pi.

Result:

open-brain/
  docker-compose.yml    # pgvector/pg16 + API, localhost-bound
  Dockerfile, .env.example
  scripts/
    models.py             # Thought: vector(768) + metadata + graph
    file_ingestor.py      # chunk 1500/200, embed, store
    embedding_service.py  # Ollama nomic-embed-text + metadata
    vector_search.py      # cosine search + capture flow
    config.py             # pydantic settings, async SQLAlchemy

$ docker compose up -d
$ python scripts/file_ingestor.py ~/notes
Ingested chunks from your files (summary/topics/people/sentiment tagged)

Telegram: what did I write about pricing experiments last winter?
-> top matches by cosine distance, answered via Ollama -> Qwen -> Claude

Use cases

• A personal knowledge base over years of markdown notes, queried from your phone.
• A clean PostgreSQL plus pgvector setup with semantic search in Python, no cloud vector DB.
• Capture-on-the-fly journaling: throw a thought at the bot, get it filed with topics and a summary.
• Cost-conscious RAG: local embeddings, local chat first, paid models only as fallback.

Known limitations

• Document RAG, not conversation memory across sessions (that is the nex-mempalace-memory-system skill); they complement each other.
• Self-hosted by design; not a cloud-managed vector database.
• Sized for a personal corpus on a Pi; a corporate-scale document set wants more hardware.
• Local Ollama is required for embeddings (one-time model pull), and the vector dimension must match the embedding model (768 for nomic-embed-text).

FAQ

What exactly is in the download?

SKILL.md, README.md, scripts/models.py, file_ingestor.py, embedding_service.py, vector_search.py, config.py, assets/docker-compose.yml, assets/Dockerfile, references on architecture, pgvector and ingestion, manifest.json, CHANGELOG.md, LICENSE.md.

Which agents does it work with?

Claude Code, Claude Cowork, Cursor, Codex CLI, and any agent that reads SKILL.md skills.

Do I need anything else?

Docker, a local Ollama install for embeddings, and a host (Raspberry Pi 5 or any Docker machine). Chat keys (Qwen, Claude) and a Telegram bot token are optional; without them it still embeds and searches.

Can I use it without the Telegram bot?

Yes. The API-only setup delivers the pgvector stack, ingestion and search; the Telegram layer is an optional interface on top.