Production RAG breaks on three things: bad chunking, missing query rewriting, and an eval set of ten queries written by the founder. We build the other 80% of the system that makes retrieval actually reliable on your real corpus: messy PDFs, stale wikis, domain jargon, multi-hop questions, and the data that updates while your agent is still pulling yesterday's snapshot.
Pixelfield is for CTOs, Heads of Product and VPs of Engineering at companies that need their AI to reason over their own data, not generic training data. The seniors who scope the work also write the code. 50+ AI features in production, fastest deployment two months, and a reputation for telling you when fine-tuning is the more expensive mistake.
The 80% past the vector searchChunking, reranking, hybrid retrievalEvaluation as CI/CD gate, not vibesFreshness SLAs and drift monitoring
Every RAG demo works on a handful of clean docs. Production has millions of pages, scanned PDFs, stale wikis, near-duplicates, multi-hop questions and users who ask in ways your test set never anticipated. We design for that reality from week one: corpus profiling, real query collection, eval set built from production data, fail-fast on the patterns that usually kill RAG.
Vector search is the famous 20% of a RAG system. The 80% that decides whether it ships is chunking strategy, query rewriting, hybrid retrieval, reranking, citation tracking, freshness SLAs, evaluation pipelines and monitoring. Most agencies sell the 20%. We build the other 80%.
An eval set of ten questions written by the founder is the most common RAG failure pattern we see. We build the eval set from real production queries, expand it with adversarial cases (multi-hop, near-duplicate, stale-data traps) and run it as a CI/CD gate. RAGAS or DeepEval, gating prompt and embedding-model changes. If you can't measure it, you can't ship it.
Around 51% of enterprise AI ships as RAG. RAG gives 80% of the value with 20% of the operational overhead. Fine-tuning is the right call for narrow, stable, high-volume behavioural change (tone, output format) where prompt engineering can't enforce it reliably. We've recommended classical RAG with better chunking after teams burned two months on fine-tuning. We'll do the same for you if it's the right call.
We sample your real corpus, classify the data shapes (structured, semi-structured, scanned, conversational), profile query patterns from real users where possible, and identify what should be deleted, restructured, or excluded before retrieval. Discovery includes an evaluation baseline on your data using a small test eval set, so you know where you're starting from.
You receive: corpus profile, retrieval architecture (standard vs Graph vs Agentic, with rationale), ingestion plan, evaluation framework, fixed-scope build quote and projected monthly run rate.
Build runs as a small senior team led by Michal Vavra. We build the ingestion pipeline (chunking per data type, OCR / table extraction where needed, deduplication, freshness CDC), the retrieval layer (hybrid BM25 plus vector plus optional graph) and the reranking pass.
Each iteration is gated on a fixed eval set sliced by query type and data shape. Retrieval precision and faithfulness are tracked from day one, not after launch.
We expand the eval set from real production queries plus adversarial cases (multi-hop, near-duplicates, stale-data traps). RAGAS / DeepEval as a CI/CD gate. Cross-encoder reranking tuned for precision. Citation tracking, source lineage, fallback chains for retrieval misses.
We run the system in shadow mode against real traffic before any user sees output. The build ships when it survives the hardening pass, not when the demo works.
At launch we wire in retrieval-quality monitoring, faithfulness drift alerts, freshness lag tracking, query-distribution drift, citation-accuracy regression and a feedback loop from human review back to the eval set. RAG systems age. We size the retainer to keep yours from quietly degrading.
Monthly retainer for monitoring, prompt and chunking tuning, embedding-model updates, on-call. Take it in-house whenever you're ready.
Direct answers to the questions CTOs and Heads of Product ask us on every <strong>scoping call</strong>.
A LangChain tutorial gets you from zero to a working demo on ten clean documents. Production RAG is a data platform problem: chunking strategy per data type, query rewriting, hybrid retrieval (BM25 plus vector plus optional graph), cross-encoder reranking, citation tracking, freshness SLAs, RAGAS / DeepEval as a CI/CD gate, drift monitoring and fallback chains. We build the 80% past the demo. If your use case fits in a tutorial, wire it up yourselves and skip us.
Around 51% of enterprise AI ships as RAG. RAG gives 80% of the value with 20% of the operational overhead. Fine-tuning is the right call for narrow, stable, high-volume behavioural change (tone, output format) where prompt engineering can't enforce it reliably. We've seen teams burn two months on fine-tuning when better chunking plus hybrid search would have got them there in two weeks. Ship RAG first; we'll tell you in discovery if your use case is one of the rare ones where fine-tuning wins.
Usually, yes. The ingestion pipeline is half the work: OCR for scanned PDFs (with handwriting where the document type needs it), structure-aware table extraction, deduplication, source-of-truth precedence across systems holding the same entity differently, freshness flags for stale wiki pages. We profile your corpus in discovery and tell you honestly when the answer is 'fix the data layer first, then build RAG'. We've recommended that more than once.
Standard RAG with hybrid search handles single-hop semantic queries on most corpora. Graph RAG when the data is genuinely relational and queries are multi-hop: org charts, invoice chains, code dependencies, compliance lineage, 'who reports to the person who approved invoice X'. Agentic RAG when the system needs to decide how to retrieve (vector vs graph vs SQL vs API). The complexity cost is real. We make the call with you in writing during discovery.
Hit rate is a vanity metric. What matters: retrieval precision, answer faithfulness, citation accuracy, context relevance. We use RAGAS and DeepEval, build the eval set from real production queries plus adversarial cases (multi-hop, near-duplicate, stale-data traps), and gate prompt and embedding-model changes on regression. RAG ships when it survives the eval gate, not when the demo works.
Discovery: fixed-fee from £2K. Small focused RAG (single corpus, single retrieval mode, embedded into one product surface): from £10K. Mid-market production builds with hybrid retrieval, reranking, evaluation pipelines and Graph or Agentic components typically land £25K-£100K. Run rate is flat monthly plus inference and embedding cost, which we engineer to minimise from day one.
Yes. Standard contract is full IP ownership: source code, prompts, prompt registry, ingestion pipelines, eval datasets, infrastructure-as-code, monitoring dashboards, runbooks. No rented layer. We document so it can be handed to your team or another vendor at any point. The only reason you stay is because the work is good.
