Establishing Link...
SecureEncrypted
Demo Active
Live demo — Sovereign Search v1.7.7
Local-First Vector Search & Privacy-Preserving Discovery
Sovereign Search Engine
Hybrid BM25 + vector search with reciprocal-rank fusion, built for product catalogues. GDPR Art. 17/20/22 endpoints, a SHA-256-chained audit log with immutable S3 archive, and a consent-driven neutral mode. Local-first by design — deploys on your cloud, your cluster, or ours. The current prototype runs on AWS EKS (us-east-1).
v1.7.7 · AWS EKS · live at search.adraca.io · P50 76 ms / P99 160 ms measured
Interface & Capabilities
GDPR consent banner on first visit
Hybrid mode results with RRF score bars
Neutral mode — privacy-only dense search
GDPR Art. 22 explanation panel
Autocomplete with structured suggestions
OpenAPI docs — full API surface
Hybrid retrieval
BM25 sparse retrieval fused with Qdrant HNSW dense retrieval via reciprocal-rank fusion (k=60), with SymSpell typo tolerance before the sparse scan.
Explainable ranking (GDPR Art. 22)
Per-result score breakdown — dense rank, sparse rank, fused RRF score — returned by /v1/explain for every search trace.
Consent-driven neutral mode
Without consent, search runs dense-only: no query rewrite, no BM25 signals, no personalisation state stored.
Chained audit log
Every search event is SHA-256-chained to the previous entry and mirrored to S3 with 365-day object-lock retention.
Architecture
Embedding
Local ONNX embedding model, SHA-256 integrity-checked at startup.
Retrieval
Qdrant HNSW dense retrieval alongside per-tenant BM25 sparse indexes.
Fusion & merchandising
Reciprocal-rank fusion of both result lists, then boost/bury/pin merchandising rules applied per tenant.
Query rewriting & reranking
Via an internal LiteLLM gateway to hosted models; PII is scrubbed before any LLM call.
Stage 01
Scrub & Correct
PII patterns (email, phone, IBAN, IP, names) removed from the query; SymSpell typo correction applied.
Stage 02
Query Rewrite
LLM rewrite through the internal LiteLLM gateway, capped at 400 ms with passthrough fallback.
Stage 03
Dual Retrieval
BM25 sparse scan and Qdrant HNSW dense lookup run against the tenant's catalogue.
Stage 04
Fuse & Explain
RRF fusion, merchandising rules, and a stored explain trace for GDPR Art. 22.
Deployment Specifications
Core Architecture
Hybrid BM25 + vector (RRF)
Embedding
Local ONNX
LLM Gateway
LiteLLM (hosted models, PII-scrubbed)
Measured Latency
P50 76 ms / P99 160 ms
Hosting
AWS EKS
adraca --log-stream
[--:--:--]SCRUBBING PII FROM QUERY...
[--:--:--]REWRITING QUERY [LITELLM · 400MS CAP]...
[--:--:--]RUNNING BM25 SPARSE SCAN...
[--:--:--]RUNNING QDRANT HNSW DENSE LOOKUP...
[--:--:--]FUSING RESULTS [RRF K=60]...
[--:--:--]APPLYING MERCHANDISING RULES [BOOST/BURY/PIN]...
[--:--:--]EXPLAIN TRACE STORED [GDPR ART.22].