Engramr

Local-first omni-search for my notes (a deliberately scoped Go project).

Engramr is a personal project to build a fast, local-first vector-based search tool over my own Markdown notes and drafts. The initial goal is "omni-search": ingest -> search -> resurface relevant passages with citations.

Later (only if useful), it may evolve into a local-first cognitive helper (summaries, synthesis, write-back), but that is explicitly not the v0 promise.

This is not a startup and not a generic RAG/chat demo.

What it is (v0)

• Local-first semantic/vector search over Markdown files
• Qdrant as local vector store
• Ollama for local embeddings
• CLI-first and intentionally single-user
• Practical and reproducible (docker compose)

What it is NOT (by design)

• No cloud services
• No accounts/auth/sync/collaboration
• No agent framework
• No "chat with your docs" product positioning
• No guarantees of stability (early project)

Status

Early (v0). Public by design: correctness and learning over polish.

Quickstart

1) Start dependencies

docker compose up -d

This starts:

• Qdrant at http://localhost:6333
• Ollama at http://localhost:11434

Pull an embedding model once:

docker compose exec ollama ollama pull nomic-embed-text

2) Build the CLI

go build -o engramr ./cmd/engramr

3) Initialize (creates collection if missing)

./engramr init

4) Ingest notes (recursively ingests *.md)

./engramr ingest path ./notes

Or ingest a web page:

./engramr ingest url https://example.com/article

URL ingest uses readability-based extraction to extract main content from web pages.

5) Query

./engramr query "Some anomaly" --top-k 10

Expected output (example):

1. notes/book/ch1.md :: Act I > Chapter 1 score: 0.73 ...snippet...

Design notes (v0)

• Markdown-aware chunking (split by headings, then paragraphs)
• Deterministic chunk and point IDs for idempotent re-ingest
• Local manifest to track indexed state per file
• Everything runs locally

Indexing semantics

Engramr uses deterministic point IDs and a local manifest file to ensure idempotent ingest operations and clean index maintenance. See docs/indexing.md for details.

Troubleshooting

See docs/troubleshooting.md for common issues and solutions.

Roadmap (non-binding)

• JSON output mode
• Optional summarization of retrieved passages
• Write-back / synthesis notes
• Desktop UI (maybe Tauri later)
• PDF ingestion

License

MIT