Case Study: Offline Semantic Search with Local Embeddings

Problem: Semantic search is a powerful tool for exploring a knowledge graph, but until recently it required a Hugging Face API token. This created a barrier for users who just wanted to try CKS, added a dependency on an external service, and introduced a common failure mode — the token was set in the user’s shell but not inherited by the MCP server process launched by Claude Desktop, leading to silent fallback to non-semantic stub embeddings.

CKS solution: The new FastEmbedEmbeddingClient (backed by the fastembed library and ONNX Runtime) runs a small, high-quality sentence-transformers model entirely locally — no API token, no network calls after the initial model download, and no external dependency. It is now the default embedding provider in cks-mcp. Combined with the existing background outbox worker and NumPy-vectorised search_embeddings, semantic search works out of the box with zero configuration.

Scenario

We built a simple graph with three objects:

Apple (Fruit)
Banana (Fruit)
Car (Vehicle)

We then ran a semantic search with the query "fruit" — without providing any explicit seed IDs — to see whether the local embeddings could correctly separate the fruits from the vehicle.

Tools Used

validate_knowledge — create the session and commit the initial version
search_semantic — semantic search without seed_ids
get_metrics — verify telemetry tracking

What Happened

1. Graph created and validated

validate_knowledge(json_data)
→ valid: true, session_id: "5245c363…", version_id: "…"

Three objects were committed: apple (Fruit), banana (Fruit), car (Vehicle).

2. Background embedding generation

The OutboxEmbeddingWorker detected the new version, extracted text representations for each object, and generated embeddings using the local all-MiniLM-L6-v2 model. This took approximately 10–13 seconds on first run (model warm-up).

3. Semantic search correctly separated fruits from vehicle

search_semantic(session_id, query="fruit")
→ matched_seeds: ["apple", "banana", "car"]
  scores: {apple: 0.705, banana: 0.671, car: 0.214}

The cosine similarity scores tell the story clearly:

Apple: 0.705 — strongly matches “fruit”
Banana: 0.671 — strongly matches “fruit”
Car: 0.214 — correctly ranked lowest, semantically unrelated

All three objects were returned only because the graph contained exactly 3 objects and the default top_k=3 returned everything. The scores themselves are what demonstrate correct semantic ranking.

4. Telemetry confirmed the tool calls

get_metrics() → tool_telemetry.tools.search_semantic
  calls: 2, success_rate: 0.5

Two calls were tracked: the first returned not_found (embeddings hadn’t been generated yet), the second succeeded. This is normal behaviour immediately after session creation.

Key Takeaways

Local embeddings work without any API keys. The server downloaded a ~90 MB model on first use and ran fully offline from that point on.
Semantic quality is excellent. A simple query like "fruit" clearly separated fruits (0.67–0.70) from a vehicle (0.21).
No configuration needed. fastembed is the default provider — no environment variables, no API tokens, no setup.
The outbox worker needs a few seconds. Embedding generation is asynchronous; a short wait after creating a session is normal.
Telemetry tracks everything. tool_telemetry in get_metrics provides per-tool call counts, success rates, and latency percentiles for observability.

Reproduce It Yourself

Install cks-mcp (v1.18.0+) and connect it to Claude Desktop.
The server will automatically use fastembed — no API keys required.
Start a chat and say:

Use cks‑mcp to create a small graph with a few objects of different categories (e.g., fruits and vehicles). Then run search_semantic with a query that clearly belongs to one category and show me the similarity scores.
Observe the cosine similarity scores — objects matching the query category will have scores > 0.6, while unrelated objects will be near 0.2 or below.

No API tokens. No cloud services. Just a conversation.