RAG Platform Development
Grounded retrieval on your own corpora, evaluated for accuracy and built to keep your data isolated.
A retrieval prototype takes a weekend. Getting it to hold up in production is the actual work, and it is mostly retrieval quality rather than model choice. Naive vector search over badly chunked documents returns plausible passages that do not answer the question, and the model then writes a confident answer on top of them.
The second problem is that quality regressions are silent. Nothing errors. The system returns worse answers, and unless retrieval quality is measured continuously, the first signal is a user who stopped trusting it.
This engagement builds retrieval that is measured, isolated per tenant, and cost-bounded.
What the engagement includes
Corpus and chunking design
How your documents are parsed, split, and enriched with metadata. Chunking strategy is the most common cause of poor retrieval and the least frequently revisited.
Hybrid search and re-ranking
Dense and lexical retrieval combined, with a re-ranking stage — because pure vector similarity misses exact-match terms that matter in technical and legal corpora.
Evaluation as a CI gate
A labelled evaluation set and retrieval metrics that run in the pipeline, so a change that degrades answer quality fails the build instead of shipping quietly.
Tenant and permission isolation
Retrieval scoped at the data layer so a user cannot surface a document they could not otherwise open. Enforced in the query, never in the prompt.
Cost and latency controls
Caching, model routing by task, and context budgeting, so the unit cost per query is known and bounded before volume arrives.
Who this is for
- Teams whose RAG prototype answers well in demos and poorly in practice
- Multi-tenant products where retrieval must respect existing permissions
- Organisations with large internal corpora and no measurement of retrieval quality
When this is the wrong engagement: If the corpus does not exist yet, or the answers require reasoning the documents do not contain, retrieval is the wrong tool and we will say so during discovery.
How it runs
- Discovery (30 minutes). A working call to map your stack, constraints, and the highest-value first step. No pitch deck.
- Scoped fixed-price proposal. A written statement of work with deliverables, milestones, and a fixed price — not an open-ended time-and-materials meter.
- Delivery. Senior architects execute against milestones, with governance applied to every AI-agent action from day one.
- Operate or hand off. Monitored operation, or a clean handoff with runbooks so your team can run it without us.
Questions we get asked
Do we need to fine-tune a model?
Usually not, and it is the most common premature optimisation in this space. Retrieval quality, chunking, and re-ranking account for most of the achievable improvement, and they are cheaper to iterate on and easier to reverse. Fine-tuning is worth considering for format and tone consistency, or a genuinely specialised domain vocabulary, and it is a poor substitute for grounding in current documents.
Where does our data go?
That is a design decision made explicitly during scoping and written down, not a default inherited from a vendor SDK. Self-hosted embedding and inference are viable where policy requires that nothing leaves your boundary. Where a hosted provider is acceptable, the architecture records exactly what is sent, what is retained, and under what terms, so a security reviewer can verify it rather than take our word for it.
How do you measure whether retrieval is any good?
With a labelled evaluation set built from real questions your users ask, scored on retrieval metrics and answer quality. This runs as a CI gate rather than a one-off report. The important property is that a change which degrades quality fails the pipeline, because RAG regressions do not raise exceptions and are otherwise found by users.
Can this run against our existing vector database?
Yes. The design works with the store you already run, and where you have not chosen one, the recommendation is based on your isolation, scale, and operational constraints rather than a preference. Migrating stores is disruptive and rarely the highest-value change available, so we only recommend it when the current one genuinely blocks a requirement.
Related services
LLMOps
CI/CD, evaluation, observability, and cost control so models ship and stay reliable in production.
Enterprise AI Architecture
Reference architectures for agents, RAG, and MLOps that hold up under real load, security review, and audit.
AI Security Consulting
Threat modeling for LLMs, data pipelines, and model endpoints — plus the controls to close the gaps.
Start with a 30-minute discovery call
No pitch deck. We map your constraints and tell you the highest-value first step — including when that step is not an engagement with us.
Book a discovery call