§ 03 · Practice
Answers with sources, not vibes.
Retrieval-augmented generation done properly: your documents, databases and live APIs behind an agent that plans, retrieves, checks its own work and cites what it used. Runs on your infrastructure or entirely offline.
§ 03.1 / Scope
Capabilities.
Ingestion
PDFs, ELN and LIMS exports, internal reports, sequence and variant databases, wikis and tickets. Chunking and metadata that match how scientists search.
Hybrid retrieval
Dense embeddings plus BM25 with reranking, tuned on your own query logs. Domain embeddings for gene, variant and compound names.
Agentic loops
The agent decides when to query PubMed, Ensembl, UniProt, ClinVar or your SQL, and when to stop. Multi-step questions get multi-step plans.
Grounded answers
Every claim links to a source span. Unsupported claims are flagged, not smoothed over.
Evaluation
Retrieval recall, faithfulness and answer quality measured on a test set built with your team, before and after every change.
Deployment
Containerised service with an API, access control and audit logs. Cloud, on-prem or fully offline with an on-device model.
§ 03.2 / Output
Deliverables.
- —A retrieval service and API you own
- —An evaluation suite with baseline numbers
- —Ingestion pipelines that keep the index fresh
- —An admin interface for sources and permissions
Tools & Standards.
nf-core and community standards where they exist, custom software where they do not.
- PostgreSQL + pgvector
- Qdrant
- LanceDB
- BM25
- LangGraph
- Claude
- MLX
- Ollama
- MCP
- Ragas
§ 03.3 / Questions
Frequently asked questions.
- Which language model does it use?
- Whichever fits your constraints. Frontier models via API for the hardest questions, or a fine-tuned small model on your own hardware when data cannot leave the building.
- How do you stop it making things up?
- Retrieval-grounded prompting, citation checks on every answer, and an evaluation set that we run before each release. Where the system is unsure, it says so.
- Can it query our databases directly?
- Yes. We expose your SQL, LIMS or internal APIs as tools with read-only credentials and log every call.
§ Practice Scoping · 03
Scope a project in agentic rag systems.
Tell us about your assay types, dataset scale, and timelines. We reply with a preliminary scoping memo within two working days.
Scope Agentic RAG Systems →