Your Knowledge Graph Needs a Search Engine

2026-03-07 · George Moon

Lattice gives you traceability — a versioned graph from research to code. You can traverse it, check for drift, and give agents full context for any requirement.

But traceability assumes you know what you're looking for. You know the requirement ID. You know which thesis to check. You can formulate the exact query.

What about the question you *can't* quite formulate? "Something about token expiry that came up in that security review." "The research that made us rethink session handling." That's not a graph traversal — it's a search problem.

Two kinds of search

Lattice search is structured. Query by ID, priority, category, or graph proximity. It's exact and fast:

lattice search --query "auth" --priority P0
lattice search --related-to REQ-AUTH-001

But some questions are fuzzy. They're about concepts, associations, half-remembered context. Structured search can't help when you don't have the right keywords or don't know which node to start from.

That's where semantic search comes in.

QMD: local hybrid search

[QMD](https://github.com/tobi/qmd) is a local search engine by Tobi Lutke. It combines three retrieval strategies — **BM25 full-text search**, **vector embeddings**, and **LLM re-ranking** — all running on-device. No API calls, no cloud dependency.

It ships with an MCP server, so agents can query it directly. 13k+ GitHub stars and actively maintained.

The key insight: QMD indexes files using glob masks. Point it at any directory and it indexes what's there. Including YAML.

Zero-config integration

QMD can index your `.lattice/` directory directly — no export step, no format conversion:

# Add your lattice as a QMD collection
qmd collection add .lattice/ --name lattice --mask "**/*.yaml"

# Annotate each layer for better context qmd context add qmd://lattice/sources "Research backing strategic theses" qmd context add qmd://lattice/theses "Strategic claims derived from research" qmd context add qmd://lattice/requirements "Testable specifications" qmd context add qmd://lattice/implementations "Code bindings satisfying requirements"

# Build the index qmd embed ```

That's it. Your knowledge graph is now semantically searchable.

How it fits together

![Lattice + QMD architecture: structured and semantic search converging in an agent workflow](/blog/qmd-lattice-integration.svg)

Two paths from the same source of truth. Lattice provides structured traversal — by ID, metadata, and graph edges. QMD provides semantic discovery — fuzzy matching, conceptual similarity, cross-layer connections you didn't explicitly model.

Complementary, not competing

These tools answer different questions from the same data:

# Structured: exact, filtered
lattice search --query "auth" --priority P0

# Semantic: fuzzy, conceptual qmd query "how do we handle expired tokens"

# Graph traversal: follow edges lattice search --related-to REQ-AUTH-001

# Conceptual discovery: find related ideas qmd query "security implications of session management" ```

Lattice search is for when you know what you're looking for. QMD is for when you know what you're *thinking about*.

Dual MCP for agents

Both tools expose MCP servers. An agent configured with both gets the best of each:

**lattice_search** — structured graph queries, drift detection, version-bound edges
**qmd_search / qmd_query** — semantic discovery, fuzzy matching, cross-collection results

Configure both in your `.mcp.json` and agents can choose the right tool for each sub-task. Need the exact spec for a requirement? Lattice. Need to find everything related to a concept? QMD.

What would make it even better

The integration works today with no changes to either tool. But two enhancements would make it smoother:

1. **`lattice integrate qmd`** — A convenience command that auto-configures the QMD collection, sets up context annotations per layer, and runs the initial embed. Replaces the six-line setup with one command.

2. **Title comments in YAML** — If lattice added a `# Title: ...` comment header to YAML files, QMD's title extractor could use it for better chunking and display. Currently it falls back to filename for non-markdown files.

Neither is required — but both would reduce friction.

Get started

# Install Lattice
curl -fsSL https://forkzero.ai/lattice/install.sh | sh

# Install QMD npm install -g @tobilu/qmd

# Initialize and index lattice init qmd collection add .lattice/ --name lattice --mask "**/*.yaml" qmd embed ```

[Lattice on GitHub](https://github.com/forkzero/lattice)
[QMD on GitHub](https://github.com/tobi/qmd)
[Live dashboard](https://forkzero.ai/reader?url=https://forkzero.github.io/lattice/lattice-data.json)