Documents & RAG¶

Jarvis indexes your personal documents (Obsidian, Notion, Drive, email, local files, PDFs) and lets you query them in natural language, in Italian and English.

What you can do¶

• 🔎 Semantic search over your notes and documents
• ❓ Natural language questions on your content
• 🔗 Citations always present (you know where you read it)
• 🌍 Multilingual: ask in English, find documents in Italian (and vice versa)
• 🖼️ Visual RAG on PDFs with figures, complex layouts, scans
• ⏱️ Incremental sync with file watchers (real-time)

Recommended stack¶

RAG framework¶

Recommendation: R2R as backend + Khoj as application layer for UI/Obsidian sync.

Embedding model 2026 (multilingual IT+EN)¶

Clear winner: BGE-M3 — supports 100+ languages in a single semantic space. Cross-lingual: ask in Italian, find docs in English.

docker compose exec ollama ollama pull bge-m3

Vector store for <1M documents¶

Jarvis default: Qdrant (reachable at qdrant:6333 in compose).

Visual RAG on PDFs¶

• ColPali / ColQwen2 — direct embedding of page images, no OCR, ideal for PDFs with complex graphics, presentations, scans
• Uses Qwen2-VL-2B as backbone

Document watchers and ingestion¶

Jarvis RAG architecture¶

flowchart TB
    subgraph Sources
        OBS[Obsidian]
        NOTION[Notion]
        DRIVE[Google Drive]
        LOCAL[Local files]
        MAIL[Apple Mail]
    end
    Sources --> WATCH[Watchers\nwatchdog · webhook · polling]
    WATCH --> PARSE[Parser\nMarkdown · PDF · DOCX]
    PARSE --> CHUNK[Chunker\nsemantic + size-based]
    CHUNK --> EMBED[Embedding\nBGE-M3 multilingual]
    EMBED --> STORE[(Qdrant)]
    STORE --> RETRIEVE[Retriever\nhybrid + rerank]
    RETRIEVE --> LLM[LLM]
    LLM --> ANS[Answer + citations]

Configuration¶

# Embedding (BGE-M3 recommended for IT+EN)
EMBEDDING_MODEL=ollama/bge-m3

# Vector store
QDRANT_URL=http://qdrant:6333

# RAG framework
RAG_BACKEND=r2r              # r2r | llamaindex | khoj
R2R_URL=http://r2r:7272

# Sources
OBSIDIAN_VAULT_PATH=/vaults/personal
NOTION_TOKEN=secret_...
GOOGLE_DRIVE_CREDENTIALS=/data/google-credentials.json
DROPBOX_ACCESS_TOKEN=...
WATCHED_FOLDERS=/data/documents,/data/research

Connect a source:

# Obsidian
docker compose exec server jarvis rag connect obsidian \
  --vault /vaults/personal

# Notion
docker compose exec server jarvis rag connect notion --token=$NOTION_TOKEN

# Local files
docker compose exec server jarvis rag connect local \
  --path=/data/documents --watch

# Google Drive
docker compose exec server jarvis rag connect google-drive

Usage examples¶

Questions about your documents¶

"What did we decide about the new backend architecture?"

Jarvis: In your March 12 notes (Obsidian → work/architecture.md)
        you wrote: "Going with append-only event-sourcing,
        snapshot every 100 events, async projections". Decision taken
        after discussion with Marco.

Cross-lingual¶

"Find that document where I talked about prompt caching"

Jarvis: Found a note in English: "anthropic-prompt-caching-cost-saving.md"
        from April 3. Want a summary?

Visual RAG¶

"That PDF with the user flow diagram, show it to me"

ColQwen2 finds the page even if the "diagram" is not OCR-readable text.

Evaluation with Ragas¶

docker compose exec server jarvis rag eval --suite=ragas-default

Measures:

• Faithfulness — is the answer grounded in the documents?
• Answer relevancy — is the answer relevant?
• Context recall — did retrieval bring back the right passages?

Privacy¶

• ✅ The whole pipeline can run 100% locally (Qdrant + Ollama + BGE-M3)
• ❌ Cloud LLM: retrieved passages are sent to the provider for the answer
• 🔐 At-rest encryption of qdrant_data volume
• 🪪 Notion/Drive tokens in a separate vault

Roadmap¶