SERVICES · RAG & KNOWLEDGE

Retrieval that cites its sources.

Production RAG systems over millions of documents. Hybrid search, reranking, cited answers, and fresh indexes. Evaluated end-to-end.

18MDOCS INDEXED FOR LARGEST CLIENT
94%ANSWER PRECISION @ TOP-3
180msP50 RETRIEVAL LATENCY
100%ANSWERS CITE A SOURCE

RAG is a system, not a vector DB.

Most 'RAG' deployments are a loose pile of chunks, a default embedding model, and hope. Ours are engineered systems — semantic chunking tuned to your document type, hybrid BM25 + vector retrieval, cross-encoder reranking, and a citation layer that refuses to answer if it can't point at a source.

We instrument the whole pipeline with retrieval metrics (MRR, nDCG, recall @ k) and end-to-end answer evals. If relevance drops after an index update, your on-call gets paged before a user notices.

We've built retrieval for legal research, insurance policy lookup, engineering docs, medical guidelines, and fintech compliance. Each needed a different chunker, a different retrieval mix, and a different rubric.

Nine retrieval capabilities.

01

Semantic chunking

Structure-aware splits — by section, by clause, by Q&A pair — not naive 500-char windows.

02

Hybrid retrieval

BM25 + dense vectors + metadata filters, fused with reciprocal rank fusion.

03

Cross-encoder rerank

Second-pass rerank with Cohere or a fine-tuned reranker for precision.

04

Cited answers

Every claim in the answer links to its source span. Refuses to answer without one.

05

Freshness

Incremental ingestion, delta indexing, and TTLs per document type.

06

Multi-tenancy

Row-level tenant filters; zero leakage across customers.

07

Eval pipeline

Golden Q&A sets, retrieval + answer metrics, regression alerts.

08

Query transforms

HyDE, query decomposition, step-back prompting where it actually helps.

09

Caching

Semantic cache over queries + documents; 40–60% cost reduction at scale.

Hybrid retrieval pipeline.

Parse → chunk → enrich → index → retrieve → rerank → answer → cite. Eight stages, each measured independently.

  • →Parser library: PDF, HTML, DOCX, transcripts, Slack exports, Notion.
  • →Enrichment: metadata extraction, entity linking, synthetic Q&A augmentation.
  • →Index: Postgres + pgvector for ≤10M docs, Pinecone above that, hybrid BM25 via Elasticsearch.
  • →Retrieval: RRF over 3–4 retrievers, cross-encoder rerank top 40 → top 8.
  • →Answer: cited JSON with span offsets, confidence per claim, refusal on no-source.

Where we've deployed it.

ORBIT LEGAL
18M
Case-law corpus indexed with per-jurisdiction filters.
STRATA INSURE
94%
Policy lookup precision @ top-3 retrieval.
LUMEN HEALTH
180ms
P50 retrieval across 2M clinical guidelines.

How we'd scope this for you.

WEEK 1
Corpus audit
Sample 500 docs, benchmark 3 chunkers, pick the mix.
WEEK 2–3
Pipeline
Parse, enrich, index, retrieve. First eval harness live.
WEEK 4–6
Tune
Rerank, query transforms, citation layer. Hit target metrics.
ONGOING
Operate
Freshness SLO, eval alerts, model upgrades.

Ready to ship rag & knowledge?

We'll scope your first use case in a 30-minute call and come back with a written proposal within 72 hours.

Book a call All services