Skip to content

Embedrock

Jump to bottom
loki-bedlam edited this page Mar 21, 2026 · 1 revision

Embedrock — Bedrock Embeddings for Loki

The Problem

OpenClaw's semantic memory search expects an OpenAI-compatible /v1/embeddings endpoint. Amazon Bedrock provides embedding models (Titan, Cohere) but uses a completely different API format. There's no built-in bridge.

Without embeddings, Loki can only do keyword-based memory search — which misses semantically related content ("what did we decide about IAM" won't find a note titled "access control policy changes").

The Solution

embedrock is a lightweight Go proxy that translates OpenAI embedding API calls to Amazon Bedrock. It's a drop-in replacement — any tool expecting /v1/embeddings works without changes.

  • Zero API keys — uses your AWS credentials (EC2 instance profile)
  • Zero config for Loki — the bootstrap script installs it as a systemd daemon
  • Single binary — no dependencies, no containers, no runtime
  • Auto-update — embedrock update self-updates from GitHub releases

How Loki Uses It

Loki Memory Search → OpenClaw → /v1/embeddings → embedrock → Bedrock (Cohere Embed v4) → vector

When Loki runs memory_search, OpenClaw sends the query text to embedrock on localhost:8089. Embedrock translates it to a Bedrock InvokeModel call using Cohere Embed v4, returns the vector, and OpenClaw uses it to find semantically similar memory entries.

Supported Models

Model ID Dimensions Notes
Cohere Embed v4 cohere.embed-v4:0 1536 Recommended — best quality
Cohere Embed English v3 cohere.embed-english-v3 1024 Good, English only
Cohere Embed Multilingual v3 cohere.embed-multilingual-v3 1024 Multi-language support
Titan Embed Text V2 amazon.titan-embed-text-v2:0 1024 Default, solid baseline

Installation

The BOOTSTRAP-MEMORY-SEARCH.md essential bootstrap handles installation automatically. If you need to install manually:

# Download (ARM64 — Graviton EC2)
curl -fsSL https://github.com/inceptionstack/embedrock/releases/latest/download/embedrock-linux-arm64 \
  -o /usr/local/bin/embedrock
chmod +x /usr/local/bin/embedrock

# Install as systemd daemon with Cohere Embed v4
sudo embedrock --model cohere.embed-v4:0 install-daemon

# Verify
curl http://127.0.0.1:8089/
# {"status":"ok","model":"cohere.embed-v4:0"}

OpenClaw Configuration

In your OpenClaw config (~/.openclaw/config.yaml), the memory search provider points to embedrock:

memory:
  embeddings:
    provider: openai
    baseUrl: http://127.0.0.1:8089
    model: cohere.embed-v4:0

No API key needed — embedrock uses the EC2 instance profile for Bedrock access.

Cost

Bedrock embedding costs are minimal:

  • Cohere Embed v4: ~$0.10 per million tokens
  • A typical memory search (100-200 token query against 1000 memory entries) costs fractions of a cent
  • Daily cost for active use: < $0.01

Self-Update

embedrock update

Downloads the latest release, verifies the checksum, and replaces the binary atomically. Restarts the systemd service if running as root.

Source

github.com/inceptionstack/embedrock — Apache 2.0 licensed

Clone this wiki locally