Use case: AI-native search and retrieval

AI-NATIVE
SEARCH &
RETRIEVAL.

Your RAG pipeline has a vector database that doesn't know about text, and a text search engine that doesn't know about vectors. XERJ runs BM25 and exact vector kNN in one query tree, one cost model, one execution pass — so hybrid search is a feature, not an integration project.

SHORT ANSWER · UPDATED 2026-10-24

XERJ is the search engine for AI agents. It runs BM25 full-text and exact vector kNN in one query tree and one execution pass, so hybrid retrieval for RAG needs no second database. Agents query it over the Elasticsearch API or through xerj mcp.

THE TWO-SYSTEM PROBLEM

Pinecone + ES = double ops. Two systems, two schemas, two billing lines, two failure domains. Data sync between them is a pipeline you have to build and maintain.
Hybrid search is an orchestration hack. Call ES for BM25, call Pinecone for kNN, merge results client-side. Latency doubles, relevance suffers, and the fusion logic is yours to debug.
Vector DB lock-in is real. Pinecone is closed-source. Qdrant has WAL corruption bugs. Milvus requires external dependencies. Migrating between them means re-indexing billions of vectors.
Cost per vector adds up. Full float32 storage means paying for four bytes per dimension, per vector, at billion-scale — before you run a single query.

THE XERJ ANSWER

One engine for text + vectors. BM25 and vector kNN share one planner. A hybrid query is one POST, not two API calls stitched together.
SQ8 quantization built in. kNN scores from int8 codes at recall@10 = 0.998. It is a precision option, not a memory one: the full-precision vector is still what is resident (#392).
Semantic search at query time. Configure an embedding endpoint and XERJ auto-embeds the query — no client-side embedding step.
No data sync. Text and vectors live in the same index, the same document, the same WAL. One write path, one consistency model.

REPRODUCIBLE · ES 8.13 · demo/playbooks/SCORECARD.md

XERJ ES 8.13

Query-string query 1.35 ms 9.35 ms 6.9×

Multi-match query 1.44 ms 5.89 ms 4.1×

kNN · k=10 · 100% recall 0.78 ms 2.64 ms 3.4×

Match query 0.57 ms 1.45 ms 2.5×

P50 · 1M-DOC LLM-TELEMETRY CORPUS · FULL 91-CELL MATRIX AT xerj.org/benchmarks →

SEE IT LIVE.

The playbook walks the full recipe — schema, ingest command, queries, and the dashboard. The playground runs on seeded data; benchmarks were measured against Elasticsearch 8.13 on 2026-04-14.

OPEN THE PLAYBOOK → OPEN THE PLAYGROUND →

READY?·REQUEST ACCESS

RUN IT ON
YOUR DATA.

Send us your embedding model and a sample corpus. We'll run hybrid search benchmarks — recall@10, latency p95, memory per million vectors — and share the reproduction.

Use case: AI-native search and retrieval

THE TWO-SYSTEM PROBLEM

THE XERJ ANSWER

SEE IT LIVE.

RUN IT ONYOUR DATA.

RUN IT ON
YOUR DATA.