Embedding

Embedding Skill

Standalone embedding service for semantic search across any database.

Architecture

┌─────────────────────────────────────────┐
│         embedding service (:8602)       │
│  Model: EMBEDDING_MODEL env var         │
│  Device: auto (CPU/GPU)                 │
└───────────────────┬─────────────────────┘
                    │
    ┌───────────────┼───────────────┐
    ▼               ▼               ▼
 memory        edge-verifier    your-project
 skill         searches         ArangoDB/etc

Quick Start

# Start the service (first run loads model ~5-10s)
./run.sh serve

# Embed text (CLI)
./run.sh embed --text "your query here"

# Embed via HTTP (after service is running)
curl -X POST http://127.0.0.1:8602/embed -H "Content-Type: application/json" \
  -d '{"text": "your query here"}'

Commands

Configuration

Swapping Models

# Use a different model for this project
export EMBEDDING_MODEL="nomic-ai/nomic-embed-text-v1"
./run.sh serve

# Or for GPU-accelerated
export EMBEDDING_MODEL="intfloat/e5-large-v2"
export EMBEDDING_DEVICE="cuda"
./run.sh serve

API Endpoints

POST /embed

Embed single text.

{"text": "query to embed"}
→ {"vector": [0.1, 0.2, ...], "model": "all-MiniLM-L6-v2", "dimensions": 384}

POST /embed/batch

Embed multiple texts.

{"texts": ["query 1", "query 2"]}
→ {"vectors": [[...], [...]], "model": "...", "count": 2}

GET /info

Service status and configuration.

{
  "model": "all-MiniLM-L6-v2",
  "device": "cuda",
  "dimensions": 384,
  "status": "ready"
}

Integration Examples

ArangoDB Semantic Search

import httpx

# Get embedding
resp = httpx.post("http://127.0.0.1:8602/embed", json={"text": "find similar docs"})
vector = resp.json()["vector"]

# Use in AQL query
aql = """
FOR doc IN my_collection
  LET score = COSINE_SIMILARITY(doc.embedding, @vector)
  FILTER score > 0.7
  SORT score DESC
  RETURN doc
"""

From Memory Skill

Memory skill can consume this service by setting:

export EMBEDDING_SERVICE_URL="http://127.0.0.1:8602"

Cold Start

First invocation loads the model (~5-10 seconds). After that, embeddings are millisecond-latency. The service logs progress:

[embedding] Loading model: all-MiniLM-L6-v2...
[embedding] Model loaded in 6.2s
[embedding] Service ready on http://127.0.0.1:8602