04 · AI Implementation
RAG & Embedding
Production-grade retrieval — measured against your golden set before it ships.
Overview
What you get
Chunking, hybrid retrieval, reranking, citation grounding, and CI eval gates so your RAG system improves with every deploy instead of silently regressing.
The problem
Why teams call us
- Hallucination rate is unknown and getting worse with each prompt tweak.
- Retrieval precision regresses silently — nobody finds out until a customer does.
- Cost-per-query is creeping up and there's no budget instrumentation.
Approach
How we work
- Start with the eval set, not the prompt. Measure first, ship second.
- Three chunking strategies benchmarked head-to-head against your data.
- Hybrid (BM25 + dense) retrieval with a reranker tuned to your domain.
- CI gates so regressions on hallucination or precision block the deploy.
Process
Week by week.
- 01 · Week 1
Eval set
Build 500+ golden query/answer set with LLM-as-judge rubrics.
- 02 · Week 2
Retrieval rebuild
Chunking benchmark, hybrid retrieval, reranker.
- 03 · Week 3–4
Grounding
Citation system, refusal logic, before/after measurement.
- 04 · Week 5
CI + handoff
Eval CI gates, dashboards, your-team-can-run-this runbook.
- Live RAG system shipped to paying customers
- Willingness to instrument production with our eval harness
- Single decision-maker on your side
- Pre-product 'we want to do RAG' exploration
- No tolerance for instrumenting production with eval traces
- Looking for a chatbot wrapper, not a measured retrieval system
Deliverables
Everything we ship
- 01Chunking redesign with three strategies benchmarked
- 02Hybrid BM25 + dense retrieval with reranker
- 03Citation system with source-grounding checks
- 04Golden eval dataset of 500+ queries with LLM-as-judge scoring
- 05CI gates so regressions block deploys
- 06Cost-per-query + latency dashboards
Outcomes
What you walk away with.
Named before/after metrics on hallucination rate, retrieval precision @5, P95 latency, and cost-per-query — methodology your team can re-run forever.
Tooling
Stack we ship against
Model- and infra-agnostic. We adapt to your stack, not the other way around.
FAQ
Real questions, technically answered.
- Will this work with our existing vector DB?
- Yes. We are infrastructure-agnostic and ship against Pinecone, Weaviate, pgvector, Turbopuffer, and Mongo Atlas.
- Do we have to switch LLM providers?
- No. We benchmark across providers and recommend, but the decision (and migration) is yours.
- What if our golden set is too small?
- We co-author it during week 1 using real production traces. 500 queries is the floor, not the ceiling.
Related engagements
Often paired with.
02 · AI Implementation
AI Copilots
In-product assistants grounded in your customers' data — that ship, not demo.
05 · AI Implementation
Fine-Tuning & Inference
Specialised models that beat general-purpose ones on accuracy and unit cost.
09 · AI Infrastructure
Production Telemetry
When your AI gets worse, your team knows in minutes — not quarters.
Next step
Ready to scope RAG & Embedding?
Book a discovery call. We'll confirm fit, sequence the engagement, and have a Statement of Work in your inbox within a week.
Refundable if we're not a fitWritten diagnostic in 48 hoursSession run by a founder, not a sales rep