// RAG and models
Custom RAG development for answers grounded in your own documents
NUMVOX provides custom RAG development for organisations that want AI to answer from their own policies, manuals, contracts and knowledge bases. Retrieval-augmented generation finds the relevant passages first, then asks the model to answer from them, with citations a reader can check.
01 / THE WORK
Who this is for.
General models do not know your internal documents, and pasting whole document stores into a public chatbot is a confidentiality decision, not a technical shortcut. Retrieval gives the model only the passages it needs, from sources you approve, under access rules you control.
Most of the quality in a RAG system comes from unglamorous work: cleaning documents, splitting them sensibly, writing good metadata and testing retrieval with real questions. That is where we spend our time.
A fit
- 01Support and operations teams searching long manuals or policy libraries many times a day.
- 02Professional services firms that want an internal assistant over precedents and templates.
- 03Product teams building an enterprise RAG chatbot into a customer-facing application.
Not a fit
- Anyone expecting a promise that answers will always match the source.
- Projects needing commitments about how a third-party model is trained.
02 / PROBLEM
The problem
A general model does not know your policies, and sending whole document stores to a public endpoint is also a decision about confidentiality. Simple RAG demos look impressive, then fail on real questions because retrieval returns the wrong passages.
03 / APPROACH
Our approach
We build the pipeline end to end: ingestion and cleaning, chunking tuned to your document types, embeddings, a vector store, retrieval with filters and re-ranking, and answer generation with citations. Quality is measured with an evaluation set of real questions and expected sources, so every change is judged on evidence.
04 / DELIVERY
What we deliver
- 01A retrieval pipeline over documents you are allowed to use
- 02Chunking, metadata and embedding choices written down
- 03Vector search with Pinecone, Chroma, Weaviate or Supabase
- 04Answer generation with citations to source passages
- 05An evaluation set and scoring so quality can be discussed
- 06Fine-tuning of an open model when retrieval alone is not enough
Not included
- A promise that answers will always match the source
- A promise about how a third-party model is trained
- Rewriting or correcting the source documents themselves
TECHNOLOGY
Models
Llama 3 · Mistral · GPT-4o · Claude 3.5
Vector stores
Pinecone · ChromaDB · Weaviate · Supabase Vector
Frameworks
LangChain · LlamaIndex · PyTorch · vLLM
Hosted models such as GPT-4o or Claude give strong answers quickly. Open models like Llama 3 or Mistral, served with vLLM in your cloud, keep data inside your environment. Supabase Vector or Chroma suit smaller corpora; Pinecone or Weaviate suit larger, multi-tenant ones.
// PROCESS
A visible path from first conversation to delivery
- 01
Audit the corpus
We review document types, formats, quality, duplication and access rules, and agree which sources are in scope.
- 02
Build the evaluation set
With your experts we collect real questions and the passages that should answer them.
- 03
Build retrieval
Ingestion, chunking, embeddings, filters and re-ranking are tuned until the right passages come back.
- 04
Generate and cite
Prompts produce concise answers with citations and refuse when sources do not support an answer.
- 05
Deploy and monitor
The system goes live with logging, feedback buttons and regular evaluation runs as documents change.
Custom RAG development versus fine-tuning
Retrieval is usually the right first step when the problem is knowledge: the model needs facts from your documents, and those facts change. Updating a document updates the answers, and citations show where each answer came from.
Fine-tuning helps when the problem is behaviour: a consistent format, tone or classification task the model performs poorly out of the box. Many mature systems use both, but starting with retrieval is cheaper and easier to evaluate.
Engagement models
RAG pilot
One corpus and one user group, with an evaluation set, to prove quality before scaling.
Production build
A full system with access control, monitoring, integrations and multiple document sources.
Augmented AI engineer
An engineer who joins your team to build and maintain retrieval systems.
Working hours from Lahore (UTC+5)
RAG engineers work Pakistan Standard Time (UTC+5). Evaluation reviews with subject experts in the UK, Europe or the Gulf are held live in a shared afternoon.
With US teams, experts can score answers asynchronously in a shared sheet or tool, and we meet early in the American day to agree changes.
Overlap and working arrangements by country: Germany, Saudi Arabia, United Arab Emirates, United Kingdom, United States.
Quality and security
Retrieval respects document permissions, so users only receive answers from sources they are allowed to read. Evaluation runs after each change catch regressions before users do.
Documents and embeddings can stay in your cloud. Where hosted models are used, we minimise the text sent and review provider retention settings with you.
How pricing works
Cost depends on corpus size and messiness, the number of sources and access rules, and the evaluation depth required. Running costs cover the vector store, hosting and model usage.
A pilot on one well-defined corpus gives you real quality numbers before a larger investment.
International clients can see up to 60% savings. Compared with typical US, UK and EU rates for equivalent roles. Actual savings depend on role, seniority and scope.
// PROOF
What you can check before you sign
- Based in
- Lahore, Pakistan (UTC+5, no daylight saving)
- Engagement models
- Project delivery, dedicated teams and staff augmentation
- Ways of working
- Written scope before work starts, and work visible in the agreed repository or environment
- Quality
- Review and testing planned with the scope, not added at the end
Frequently asked questions
What is retrieval-augmented generation in plain terms?
It is a two-step process. First, the system searches your documents for passages relevant to the question. Second, it gives those passages to a language model and asks it to answer using only that material, with citations. The model supplies language skills; your documents supply the facts.
How do you measure RAG quality?
We build an evaluation set of real questions with the passages that should answer them, then score two things: whether retrieval finds the right passages, and whether answers are correct and supported. Scores are tracked over time, so every change to chunking, models or prompts is judged against the same benchmark.
Can a RAG system respect document permissions?
Yes, and it should. We store access metadata with each chunk and filter retrieval by the user's permissions before anything reaches the model. A user who cannot open a document in your system should never receive an answer built from it. Permissions are tested as part of the evaluation set.
Can the whole system run inside our own cloud?
Yes. Open models such as Llama 3 or Mistral can be served with vLLM in your cloud account, alongside a self-hosted vector store, so documents and queries never leave your environment. This costs more to operate than hosted APIs, so we compare both options against your security requirements.
What kinds of documents work best?
Text-rich documents with clear structure work best: policies, manuals, procedures, contracts, product documentation and support articles. Scanned PDFs, tables and slides need extra processing but are usually workable. Very short, conflicting or outdated documents are the bigger problem, and we flag them during the corpus audit.
How much does custom RAG development cost?
It depends mainly on the size and quality of your documents, the number of sources, access rules and how rigorous the evaluation must be. A focused pilot costs far less than a multi-source enterprise system. Our guide to RAG development cost explains the drivers so you can budget before talking to vendors.
// NEXT STEP
Talk through your requirements.
Tell us the workflow, the users, and what you already know. We reply by email after we have read the brief. This form is not a place for passwords or sensitive personal data.
Project enquiry
Discuss Custom AI Model & RAG
Share the context a reviewer needs. We follow up by email once the brief is clear.
