Skip to content
AB19 Innovations

RETRIEVAL · GROUNDING · CITATION

Answers grounded in your own material.

A general model knows the internet and nothing about your business. Retrieval closes that gap — but only if the retrieval is good. Most RAG systems fail at the retrieval step, not the generation step, and that is where we spend the effort.

Cited
Every answer traceable to a source
Scoped
Permissions enforced at retrieval
Measured
Retrieval quality scored independently

THE SHORT VERSION

Why retrieval is the hard part

Generation is largely solved: given the right context, modern models write a good answer. The difficulty is deciding what the right context is. Chunking strategy, embedding choice, hybrid keyword and vector search, reranking, and knowing when to return nothing at all.

We evaluate retrieval separately from generation. If the correct passage was never fetched, no amount of prompt engineering will save the answer — and you want to know which of the two failed.

The difference is where the effort goes.

Naive RAG

  • Fixed-size chunks that split tables and headings apart
  • Pure vector search that misses exact identifiers and part numbers
  • No citations, so nobody can verify an answer
  • Every user sees every document, regardless of permissions

Engineered retrieval

  • Structure-aware chunking that preserves the document’s own hierarchy
  • Hybrid search with reranking, so exact terms and meaning both work
  • Inline citations back to the source passage
  • Permission filtering applied at query time, not after

WHAT IT INCLUDES

Scope, in plain terms.

Every engagement is shaped to the problem, but this is the ground it typically covers.

  • Knowledge assistants

    Ask-your-documents over policies, contracts, manuals and internal wikis.

  • Retrieval pipelines

    Ingestion, parsing, chunking, embedding and index maintenance.

  • Hybrid search

    Vector plus lexical retrieval with reranking for precision on real queries.

  • Structured extraction

    Turning unstructured source material into typed, validated records.

  • Content generation

    Drafting workflows with brand constraints and a human approval step.

  • Retrieval evaluation

    Recall and precision measured per query class, not averaged into meaninglessness.

What you get

  • Ingestion pipeline with incremental re-indexing
  • Hybrid retrieval service with reranking and permission filters
  • Answer synthesis with inline citations
  • Retrieval and answer evaluation suites, run in CI
  • Admin surface for inspecting what was retrieved and why

QUESTIONS

The things people ask first.

Do you fine-tune or use retrieval?

Retrieval first, almost always. It is cheaper, updates instantly when your content changes, and produces citable answers. Fine-tuning is for format and behaviour, not for facts.

How do you stop it inventing things?

Ground every claim in retrieved context, cite the source inline, and make "I do not have that" a valid, tested response rather than a failure.

Can it respect our access controls?

Yes, and it must. Permissions are applied as filters during retrieval so a user can never be shown a passage they could not otherwise open.

Often paired with

Need generative ai & rag?

Tell us the problem and the constraint. We will tell you what it takes — and whether this is even the right service for it.

Let’s talkAll services
founder@ab19innovations.com