Offers three modes (dense, sparse, hybrid), reranking, and metadata filtering, plus retrieval quality observability. The model is natively multilingual (including Polish) and handles mixed languages in a single index — a question asked in Polish finds the answer regardless of the source language. A single query (embedding + retrieval) typically drops below ~50 ms on a typical index — you measure that figure in the retrieval observability dashboard, we don't promise it in isolation from your data. Self-hosted deployment gives you control over cost and privacy. This is the engine we use for Company GPT and internal knowledge search tools. Underneath runs the BGE-M3 model — technical specs (1024-dim, embeddings); it's the same local model that powers this site's Concierge RAG.
cashcrown // search engine
BGE-M3 Search
bge-m3-searchliveA self-hosted embeddings and semantic search service with a dashboard and full retrieval observability.
<50ms
latency
3×
search modes
self-host
deploy
// 01
Problem
Keyword search does not understand intent, and off-the-shelf embedding APIs send your data outside and scale their cost with your traffic. Without your own retrieval, every RAG stack is built on someone else's foundation.
// 02
Capabilities
- search modes: dense, sparse, hybrid
- reranking and metadata filtering
- retrieval quality observability
- a foundation for any RAG stack
// 03
Delivery scope
- A local embeddings service (1024-dim) — content never leaves your infrastructure
- Dense/sparse/hybrid search + reranking + metadata filtering
- A dashboard with full retrieval observability (what was found and why)
- One-command reindex + embedding parity locally and in production
We assemble this product from measured models and components — see the full live pipeline:
// 04
Proof-led pilot
Pilot package: local retrieval and RAG
BGE-M3 is packaged as a local knowledge-search foundation, not an abstract embeddings model. The proof connects product, playground, build recipe and data residency.
1024
embedding dimensions
3x
dense, sparse, hybrid
local
index and embeddings
Evidence
Embeddings playground
The public demo shows local vectors and semantic similarity without storing data.
view proofConcierge RAG
The recipe shows how retrieval powers cited answers and human escalation.
view proofCompliance and data residency
The trust center describes GDPR boundaries, PII, registers and sensitive-data operating rules.
view proofPilot path
- 01We choose one knowledge corpus and measure retrieval quality before generation.
- 02We index locally, add metadata filters and source citations.
- 03We leave a report: relevance, knowledge gaps, operating cost and rollout decision.
// 05
Integrations
QdrantpgvectorFastAPIMeiliSearch
// 06
Architecture
Select a node to see its description and data flow.
Architecture — text description
- Ingest / data
- Data sources — Content, documents and events feed the index. (Flows to: BGE-M3)
- BGE-M3 — Local embeddings (1024-dim) — PII stays on-prem. (Flows to: OpenClaw router)
- AI core
- OpenClaw router — The only entry to the models — masks PII, picks the model, throttles. (Flows to: BGE-M3 Search)
- BGE-M3 Search — A self-hosted embeddings and semantic search service with a dashboard and full retrieval observability. (Flows to: API)
- Serve / integrations
- API — REST surface — cited answers with access control. (Flows to: Qdrant, pgvector, FastAPI, MeiliSearch)
- Qdrant — Integration: Qdrant.
- pgvector — Integration: pgvector.
- FastAPI — Integration: FastAPI.
- MeiliSearch — Integration: MeiliSearch.
// 07
Live demo
// 08
FAQ
How much does it cost and what does the rollout look like?
We work in ranges that depend on scope — we start with a fixed-cost pilot so you see the value before a larger investment. When the solution takes back a dozen to several dozen hours of work a month, it usually pays off in 2–4 months. Calculate the return in the ROI calculator.
Does it integrate with our systems and is it GDPR-compliant?
Yes. We connect to your systems through controlled, allow-listed tools, and irreversible actions require confirmation (a human-gate). We mask PII before the cloud and handle sensitive paths locally — GDPR and AI Act compliance designed in from the start.
Does it run offline / on-prem?
Yes — it is a self-hosted service; the embeddings and the index stay with you.
What is it good for?
Knowledge search, RAG, deduplication, and document or offer similarity.
// →Related