shared-context-layer

Before building more bots, I need to establish a single shared context layer that all three bots can query.

this is to avoid rebuilding the Obsidian integration three times and ensures consistent knowledge across the system.

The Obsidian vault is the canonical knowledge source -> A watched folder pipeline embeds markdown files into a shared vector store (ChromaDB or SQLite + embeddings).
Any bot queries this store via a common internal API.
No bot owns its own copy of the notes.

Implementation sequence

1. Watcher script — monitors vault folder, triggers on new/changed .md files
2. Embedding pipeline — chunks and embeds into the vector store
3. Query API — a simple FastAPI endpoint: POST /context {query, top_k}
   returns relevant note chunks
4. Supervisor Bot will use this instead of any hardcoded context
5. Confidence Bot will also read from the same store

Pipeline
your note text
↓
chunk by heading (keep chunks semantically coherent)
↓
nomic-embed-text converts each chunk to 768 numbers
↓
ChromaDB stores [numbers + original text + metadata]

query arrives (“what is my research about”)
↓
nomic-embed-text converts query to 768 numbers
↓
ChromaDB finds stored vectors closest to query vector
↓
returns original text of those chunks

When running API again in future sessions I will need two terminal tabs – one running the watcher, one running the API:

# tab 1
python3 watcher.py

# tab 2
python3 -m uvicorn api:app --reload

Key open questions

• Does the Activity Log feed PhD-Live automatically or via manual curation?
  system-open-questions

What this is not

This is not a full RAG system build. The vector store is infrastructure,
not a product. Keep it thin — the interesting decisions live in the bots,
not the plumbing.