Services · AI & Machine Learning
RAG System Development
Retrieval over your documents, with citations you can check
What this is
What rag system development covers
Retrieval-augmented generation for legal, financial, healthcare, and engineering teams whose answers must cite a source. We build the ingestion pipeline, chunking and metadata strategy, hybrid retrieval, and the answer layer with span-level citations, then measure retrieval quality on questions your experts wrote. Deployments run in your VPC or on-premise when the corpus cannot leave.
Typical timeline: 6–12 weeks
Who hires us for this
- Legal and compliance teams searching case law, contracts, or policy libraries
- Risk and finance teams querying proprietary frameworks and filings
- Engineering and operations groups with decades of manuals, tickets, and runbooks
Problems we get hired for
What usually goes wrong before we are called
The answer is fluent and wrong
Answers are grounded in retrieved spans and every claim carries a citation to the page or clause. Unsupported claims are suppressed rather than guessed.
Retrieval finds the wrong document
Hybrid search (BM25 plus dense vectors), metadata filters, and reranking are tuned against a question set your experts wrote. Recall at k is a tracked number, not a feeling.
The corpus cannot leave the building
The full stack, including embedding and generation models, runs inside your VPC or on-premise. The legal research deployment kept privileged documents on client infrastructure.
How the engagement runs
Steps and the artifact each one produces
- 01
Corpus audit
Document types, formats, access rules, and update frequency. Decide what to index and what to exclude.
Artifact: Corpus map and access matrix
- 02
Golden question set
Fifty to two hundred real questions with expected sources, written with your experts.
Artifact: Question set and retrieval scorecard
- 03
Pipeline build
Ingestion, chunking, metadata, hybrid retrieval, reranking, and cited generation. Tuned against the scorecard.
Artifact: Retrieval pipeline with quality dashboard
- 04
Handover
Re-indexing jobs, permission sync, drift alerts, and a runbook for adding new document types.
Artifact: Ops runbook and refresh schedule
Stack
Tools we reach for
Chosen per engagement against your existing platform. Nothing here is mandatory; everything here has shipped.
| Ingestion | Unstructured · Docling · OCR fallbacks · Airflow or Temporal for schedules |
|---|---|
| Retrieval | pgvector · Qdrant · Weaviate · BM25 via OpenSearch · Cohere or BGE rerankers |
| Embeddings | OpenAI text-embedding-3 · BGE-M3 · Nomic for on-premise |
| Generation | Claude · GPT-4 class · Llama 3 or Qwen in VPC |
| Quality | Ragas · Custom recall@k dashboards · Citation coverage checks |
What you receive
Deliverables
- Corpus map with access rules
- Golden question set and retrieval scorecard
- Ingestion and hybrid retrieval pipeline
- Cited answer interface or API
- Re-indexing jobs, permission sync, and runbook
Proof
Where this has run
Legal
Case Law Research Assistant for Mid-Market Law Firm
Hybrid RAG with citations: research time per matter from 4–6 hours down to ~45–90 minutes.
Research time per matter: 4-6 hours → ~45-90 minutes
Financial Services
Proprietary Risk Framework Assessment: RAG Over Fine-tuning
RAG POC ~$40K vs ~$300K fine-tuning path: same validated quality band, overnight framework updates.
Build cost: ~$300K (fine-tuning) → ~$40K (RAG POC)
Products that ship inside this service
- Xenith RAG · Hybrid retrieval over private documents, with citations you can check.
- Xenith Knowledge · Institutional knowledge captured from Slack, email, and wikis, then searchable.
- Xenith Private AI · Open-weight LLM stack inside your VPC or air-gapped network.
Engagement models
How we can be hired for this
| Model | Shape | Typical duration |
|---|---|---|
| Fixed-scope pilot | Defined workflow, defined integration points, defined exit metric. Priced as a project. | 4–12 weeks |
| Production retainer | Monthly capacity for evals, monitoring, model refreshes, and new workflows. Month-to-month. | 3+ months |
Pricing and phase gates are described on How we work.
Questions buyers ask
Frequently asked questions
- Which vector database do you use?
- pgvector when the corpus is under a few million chunks and you already run Postgres; Qdrant or Weaviate when you need dedicated scaling or filtering. Our vector database guide lists the trade-offs. Hybrid search with BM25 matters more than the database choice.
- Can RAG run fully on-premise?
- Yes. Embedding models (BGE-M3, Nomic), rerankers, and generation models (Llama 3, Qwen) all run on GPUs you own. The on-premise RAG guide covers hardware sizing; the legal research case study ran on client infrastructure.
- RAG or fine-tuning?
- RAG when the knowledge changes or must be cited; fine-tuning when the task is a narrow style or classification. The financial risk case study chose RAG over fine-tuning because the framework documents change quarterly and analysts needed citations.
- How do you handle document permissions?
- Access control lists are indexed as metadata and enforced at retrieval time, so a user only sees chunks they could open in the source system. Permission sync runs on the same schedule as re-indexing.
When this is the wrong fit
- Small, static FAQ sets; a well-written page or a search index is cheaper.
- Use cases where the answer must be computed from structured data; that is a SQL or analytics problem.
Next step
Send the workflow. We reply with an architecture sketch.
Describe the queue, the system of record, and who signs off today. First brief is free; NDA first if you prefer.