Services / RAG
RAG & retrieval systems.
Most "AI search" projects fail because the retrieval is wrong, not because the model is wrong. We build hybrid retrieval that actually returns the right documents, with reranking tuned on your data and evals that run on every commit.
When this is the right fit
- Your users search inside a corpus of documents, tickets, code, or product data.
- Your current search returns "results" but not answers.
- You're considering — or have already shipped — a vector database, and the results are mediocre.
- Compliance, support automation, internal knowledge, or product Q&A are the use cases.
What we ship
- Hybrid retrieval pipeline — BM25 + dense embeddings, with a learned reranker.
- Chunking strategy chosen for your specific document shape.
- Embedding model selection with offline evals on your data, not benchmark scores.
- Reranker — typically a cross-encoder, tuned on a labeled subset of your queries.
- Evaluation suite — Recall@k, MRR, and a small LLM-as-judge harness for end-to-end answer quality. Runs in CI.
- Production API — deployed on your infrastructure, instrumented for cost and latency.
- Operator dashboard — what queries are happening, what's failing, what's expensive.
- Runbook — how to add documents, retune the reranker, debug a bad answer.
Typical timeline
| Week | What ships |
|---|---|
| 1 | Scope, architecture, eval plan, dataset of labeled queries. |
| 2 | Baseline pipeline (chunking + dense retrieval) running on your data. |
| 3 | Hybrid retrieval + reranker; eval suite passing on a target threshold. |
| 4 | API in your staging environment; observability live. |
| 5–6 | Production deploy + handoff. |
Range: 4–6 weeks.
FAQ
We already have a vector DB. Are you replacing it?
Not necessarily. We start by evaluating what's there. If it's working, we keep it. If it's not, we tell you why before recommending a swap.
Will you use OpenAI embeddings, or open-source?
Whichever wins on evals on your data. We benchmark both during week one.
Do you handle the ingestion side?
Yes. Document loading, deduplication, and the chunking pipeline are part of the engagement. Connectors to specific systems (e.g., Salesforce, Notion) are scoped on top.