Intelligence / RAG system
RAG system development: answers from documents nobody has time to read
Thousands of contracts, reports, or manuals turned into something you can ask questions of, with every answer traced to a page. $5,997 in fourteen days, starting with a pilot on your own documents.
- Fixed price
- $5,997
- Ships in
- 14 days
- Quote in
- 48 hours
Code transfers to you. Deposit back if we miss the date.
Organisations accumulate documents faster than anyone can read them. Contracts, technical manuals, research, regulatory filings, a decade of internal decisions. The knowledge is there. Finding it takes someone who already knows where to look.
Retrieval-augmented generation solves the finding problem. Ask a question in plain language, get an answer assembled from the relevant passages, with citations to the exact document and page so you can verify it. The model never answers from memory, only from what was retrieved.
Retrieval over a private document library is the newest service on this menu, and we would rather scope it honestly than sell it hard. So every build starts with a paid pilot on a sample of your own material: your documents, your worst scans, your real questions, and you see measured results on that sample before committing to the rest. The pilot fee comes off the build.
Best for
- Professional services firms with large contract archives
- Manufacturers with technical manuals and specifications
- Compliance and legal teams working across regulation
- Consultancies mining a decade of past reports
- Any team where the answer exists but finding it takes an hour
Built with
- Anthropic Claude
- OpenAI
- Pinecone
- pgvector
- LangChain
- Next.js
- Supabase
- Vercel
What's included
Everything in the $5,997 scope
Written out in full, because a fixed price only means something when the scope is specific.
A pilot on your own documents first
Before the full build we ingest a representative sample of your material and run your real questions against it. You see how it handles your most awkward queries and your worst scans, and you decide whether to continue. The pilot fee comes off the build price if you do.
Document processing pipeline
Ingestion for PDFs, Word, spreadsheets, slides, HTML, and scans, including OCR. Layout-aware chunking that keeps tables and clauses intact rather than slicing them mid-sentence, which is where most homemade RAG systems fail.
Hybrid search and re-ranking
Vector search for meaning plus keyword search for exact terms: part numbers, clause references, names, combined and re-ranked. Semantic search alone misses precisely the queries professionals actually type.
Cited, grounded answers
Every claim links to its source document and page. Users verify in one click, which is what makes the system usable for work with real consequences.
Permission-aware retrieval
Retrieval respects who is asking. Documents are scoped by role or team so the system cannot surface material a user is not entitled to see, filtered before retrieval rather than after generation.
Search interface
A clean query interface with filters, conversation history, source previews, and the ability to scope a question to a folder, client, or date range.
Evaluation harness
A test set of real questions with known correct sources, scored on retrieval precision and answer accuracy. Re-runnable whenever content or models change, so quality is monitored rather than assumed.
What you receive
The concrete list handed over on ship day.
- A deployed retrieval system over your document library
- Ingestion pipeline with OCR for scanned material
- Hybrid vector and keyword search with re-ranking
- Cited answers linked to document and page
- Role-based permission filtering
- Web search interface with filters and history
- Admin tools for adding, removing, and re-indexing documents
- Evaluation harness with a scored baseline
- Thirty days of tuning support
What this tier does not cover
Stated up front. Anything here can be quoted as an add-on.
- A guaranteed accuracy figure before the pilot has run
- Taking actions on documents, such as editing or signing
- Real-time sync with a document management system, priced per connector
- Model fine-tuning on your corpus
- Document libraries above 100,000 files, quoted separately
- Model and embedding API usage, billed to your own account
On-time guarantee
Miss the timeline on a fixed-price build and your deposit comes back. No negotiation, no invoicing games.
How it runs
From brief to live in 14 days
The same four phases on every build. No discovery retainer, no phase two.
- 0148 hours
Scope lock
You pick a tier or send a brief. We come back with a written scope, every screen, every integration, a fixed price, and a ship date. Nothing starts until you approve it, and nothing outside it is assumed.
- 02Day 1
Asset handoff
A single onboarding form collects logins, brand assets, and access. One call if the project needs it, none if it does not. We select the stack and start the same week.
- 03The bulk of it
Build in the open
Senior delivery from day one, no junior handoff. You get a walkthrough video at every functional milestone and a live preview URL you can open at any point, so there is no reveal at the end.
- 04Launch + 30 days
Ship and hand over
It goes live on your infrastructure, in your accounts, with the repository transferred. Thirty days of support covers the bugs and questions that only appear once real users arrive.
FAQ
RAG system questions
Answered here so you can decide without a sales call.
How many documents can it handle?
Enough for a typical professional library, and the architecture scales beyond it. We size the build against your actual corpus during the pilot rather than quoting a ceiling we have not tested on your material, in practice document quality and structure matter far more than raw document count, and a thousand badly scanned contracts are harder than ten thousand clean ones.
Does it work on scanned PDFs?
Yes. OCR is part of the pipeline. Quality tracks scan quality: clean scans are near-perfect, faxed documents from 1998 less so. We test a sample of your worst material during scoping rather than discovering it later.
How accurate is it?
We will not give you a number before we have seen your documents, and you should be sceptical of anyone who does. Accuracy depends almost entirely on your material and the questions you ask of it. What we do instead is measure it during the pilot, on your corpus, and show you the result before you commit to the full build. Every answer also cites its source, so anything that matters can be checked in one click rather than trusted.
Can it keep our confidential documents confidential?
Documents live in your own vector store on your infrastructure. API traffic runs under business terms that exclude training. Permission filtering happens before retrieval, so restricted material never enters a prompt. For strict environments we can run open-weight models in your own cloud instead.
How is this different from just uploading files to ChatGPT?
Context limits, permissions, and citations. ChatGPT handles a handful of documents per conversation with no access control and no verifiable sourcing. A RAG system searches your entire library, respects who is asking, and cites the page, which is what makes it usable for work you are accountable for.
What does it cost to run monthly?
Typically $100 to $400: vector database hosting plus embedding and generation calls. Re-indexing is only triggered by new documents, so steady-state cost tracks query volume rather than library size. The pilot gives us your real corpus size and query pattern, so the projection you get afterwards is based on your numbers rather than an average.
Other AI app services
AI chatbot
A support chatbot that answers from your own material and escalates when it should.
- Price
- $1,997
- Timeline
- 10 days
AI agent
An agent that finishes a real process across your tools, not one that answers questions about it.
- Price
- $4,997
- Timeline
- 14 days
Start a rag system build
$5,997, shipped in 14 days. Send a brief and we will confirm scope and a start date within 48 hours.
No retainer, no hourly billing. Quote in 48 hours.