RAG development services
RAG systems that answer from your data — and show where the answer came from
We build retrieval-augmented generation (RAG) pipelines and AI knowledge bases that search your documents, databases and product content, then answer with citations. The focus is on what makes RAG work in production: ingestion quality, hybrid retrieval, access control and an evaluation set that tells you whether it is actually right.
RAG fails at retrieval, not at generation
When a RAG system gives a wrong answer, the model is rarely the cause. The right passage was never retrieved: the PDF was parsed into garbage, a table lost its headers, chunks split a definition from its exception, keyword-heavy questions missed in pure vector search, or the user was shown a document they should not have access to.
So we spend most of a RAG project on ingestion, retrieval and evaluation — and only then tune prompts. The result is a system whose accuracy you can measure and improve, instead of one you hope is right.
How we build retrieval that holds up
The pipeline decisions below are where production RAG quality is won or lost.
Ingestion that respects structure
Structure-aware parsing for PDFs, tables and scanned documents (including vision-model extraction where plain text extraction fails), with headings and section context preserved on every chunk.
Hybrid search, then rerank
Vector similarity and keyword (BM25/full-text) search combined with reciprocal rank fusion, with optional reranking of the top candidates. Part numbers, names and exact phrases stop slipping through.
Metadata filtering and tenant isolation
Every chunk carries tenant, permission and document metadata; filters run inside the query, so one customer can never retrieve another customer's data.
Answers with citations
Responses cite the passages they used, and the system says "I don't know" when retrieval comes back empty instead of inventing an answer.
Freshness
Incremental re-indexing when documents change, content-addressed deduplication, and deletes that actually remove data from the index.
Measured, not eyeballed
A question set with expected sources and answers, scored for retrieval recall and answer faithfulness on every change to chunking, embeddings or prompts.
What we build
Internal knowledge assistants
Policies, SOPs, contracts, wikis and tickets in one searchable assistant that cites its sources and respects existing permissions.
Customer-facing answer engines
Product documentation and help-centre search inside your app or site, deflecting support tickets with answers users can verify.
RAG inside your SaaS
Per-customer knowledge bases in a multi-tenant product: each tenant uploads its own content and gets answers only from it.
Document processing pipelines
Extracting structured fields from contracts, invoices and forms, with retrieval used to ground extraction in the right section.
RAG rescue
An existing RAG prototype with disappointing answers: we build an evaluation set, find where retrieval fails and fix the pipeline.
Typical use cases
Support deflection
Answering customer questions from docs and past tickets, with links to the source article and hand-off to a person when confidence is low.
Sales and proposal support
Finding the right case study, security answer or contract clause across hundreds of documents in seconds.
Regulated knowledge
Policies and procedures where every answer must be traceable to an approved source document.
Course and catalogue search
Matching natural-language questions to structured catalogues — products, courses, listings — combining semantic search with hard filters.
Our approach
- 1
Collect real questions
We gather 50–200 questions real users ask, with the documents that should answer them. This becomes the evaluation set.
- 2
Audit the content
We check formats, structure, duplication, permissions and update frequency. Most RAG problems are visible here, before any code.
- 3
Build a baseline and measure
A simple pipeline, scored on retrieval recall and answer faithfulness, so every later change is judged against numbers.
- 4
Improve the weakest stage
Parsing, chunking, hybrid retrieval, reranking or prompting — whichever the evaluation shows is losing the most answers.
- 5
Ship with monitoring
User feedback, unanswered-question logs and periodic re-evaluation, so quality does not quietly decay as content changes.
Experience behind this service
Our founder designed and runs a multi-tenant RAG pipeline in production on an AI EdTech platform used across 20+ US schools: hierarchical chunking, hybrid semantic + keyword search with reciprocal rank fusion, per-school tenant isolation on pgvector, and vision-model PDF ingestion. The client work below shows the same engineering discipline applied to structured AI output.
AI-generated assessments and executive reports for a leadership consultancy
A team-assessment app where admins generate assessment templates from a topic with an LLM, participants answer through share links, and the system writes executive-summary reports.
Read the case studyAI integration · MCPConnecting an operations platform to Claude with a remote MCP server
A remote Model Context Protocol server that lets the consultancy's staff run their admin platform by asking Claude — invoicing, scheduling, follow-ups, time entries and contracts — with the same permissions they have in the app.
Read the case studyTechnology we use
Vector & search
- pgvector (PostgreSQL)
- PostgreSQL full-text search
- Reciprocal rank fusion
- Rust SIMD vector search (founder-built)
Embeddings
- OpenAI text-embedding-3
- Binary & scalar quantisation
Orchestration
- LangChain
- LangGraph
- Vercel AI SDK
Ingestion
- Vision-model PDF extraction (Claude)
- Hierarchical chunking
- Content-addressed deduplication
Evaluation
- Retrieval recall & faithfulness scoring
- promptfoo
- Langfuse
- LLM-as-judge with human calibration
Integration and deployment considerations
Use the database you have
If you run PostgreSQL, pgvector is often enough and avoids a new vendor — it is what we use in production. A dedicated vector database is worth its operational cost only when scale or features demand it.
Access control
We mirror the permissions of the source systems (SharePoint, Google Drive, your app's roles) at retrieval time rather than trusting the prompt to hide things.
Data residency and retention
Embeddings and documents stay in your cloud account; model calls go through enterprise or zero-retention endpoints where required.
Latency and cost
Caching, smaller models for query rewriting and reranking only the top candidates keep answers fast and per-query cost predictable.
What drives the cost
A RAG system over one well-structured document set with a single user group is a small project. Cost grows with the number of sources, how messy the documents are, permission complexity, and the accuracy bar you need to hit. We quote after looking at a sample of your content.
- Number and type of content sources (PDFs, scans, wikis, databases, tickets)
- Document quality — tables, scans, inconsistent formatting
- Permission model and multi-tenancy
- Required accuracy and size of the evaluation set
- Update frequency and re-indexing needs
- Query volume and latency targets
Who this is for
- SaaS companies adding "ask your data" features for their customers.
- Organisations with large document collections that staff struggle to search.
- Support teams wanting answer deflection that cites approved sources.
- Teams with a RAG prototype whose answers are not good enough to ship.
Further reading from our engineers
- RAG vs fine-tuning for business knowledge: a decision framework
When to retrieve, when to fine-tune, when to do both — and the retrieval failures that actually decide whether your AI answers are right.
Need an engineer inside your own team instead of a project? Hire AI developers from DablewDevs on an hourly or dedicated basis.
Related services
RAG Development: common questions
Tell us what your users need to find
Tell us what you are building or what needs fixing. You will hear back from an engineer — not a sales rep — within one business day, usually with questions or a rough approach.
Your enquiry goes straight to our engineering team — the people who will scope and build your project, not a sales desk.