rag Service

Custom RAG Development

Production-grade RAG systems over enterprise corpora — auditable, fine-tunable, EU-deployed.

The problem

Teams that tried retrieval-augmented generation with a weekend hackathon discovered between a PoC and production there are 7 architecture layers they didn't think about. Hallucinations, latency, audit, maintenance — all appear at month 3.

How it works

1
Week 1-2: corpus discovery — what documents, what format, what metadata, what typical questions. Output: architecture spec + vendor decision matrix.
2
Week 3-6: implementation of hybrid retrieval (BM25 + dense), citation grounding, query rewriting, reranker. Iterations with team on real queries.
3
Week 7-10: eval pipeline — automated (precision@k, MRR, faithfulness) + manual review on 100 queries. Iterations until agreed threshold.
4
Week 11-12: production hardening — full audit log, monitoring, runbooks, team training. Hand-off with 12 months support included.

Capabilities

Hybrid retrieval (BM25 + dense + reranker)

High recall on vague queries (BM25), precision on semantic queries (dense), final ranking with cross-encoder reranker. Pattern applied to every client.

Citation grounding on every answer

Answer includes link to exact document fragment. For sectors where source matters (legal, financial, healthcare), non-negotiable.

Query rewriting + decomposition

Multi-step questions decompose into sub-questions (HyDE, CRAG). Quality on complex queries increases 25-40% vs naive retrieval.

Automated eval pipeline

Precision@k, MRR, faithfulness, citation accuracy — measured on every release. Automatic regression detection — a new model doesn't reach production without beating baseline.

Full audit log

For every query: timestamp, user, prompt, retrieval results, ranking, final prompt to LLM, answer, citations, duration. For forensics after N months.

EU-resident infrastructure

Deployment on-premise or in EU private cloud (Romania, Frankfurt, Amsterdam). Never US/Asia — for Schrems II compliance.

Deliverables

✓ Architecture spec + decision matrix
✓ Production-grade codebase (Python + FastAPI + Postgres/Qdrant)
✓ Automated eval pipeline (CI integrated)
✓ Operational runbooks + team training
✓ 12 months post-launch support

Typical timeline

6-12 weeks end-to-end, depending on corpus size and query complexity.

FAQ

How does it compare to a Pinecone/Weaviate SaaS? +

Vector store SaaSes are one of the 7 layers. We work on all 7 — corpus ingestion, hybrid retrieval, reranking, query rewriting, citation, eval, audit. SaaS solves layer 4; what we do covers end-to-end.

Can we start with a small PoC? +

Yes. We offer 2-week Discovery Sprint with fixed cost — produces architecture spec + a functional PoC on 1000 representative documents. You decide after whether to continue to full implementation.

What LLM do you use? +

Model-agnostic. For Romanian work: fine-tuned RO-corpus models (Qwen3 family, Gemma). For complex English work: frontier models (Claude, GPT-4). Decision made with client team based on latency/cost/quality trade-off.

We start with a 30-minute conversation.

Free AI-readiness audit for companies with 50+ employees. We reply within 24 hours.

Open contact form →