AI enablement · Knowledge and retrieval
Answers from your own material, with the source attached
Ingestion, retrieval quality, grounding and citation, so a question asked against your own documents returns something a person can verify. Retrieval quality is the whole problem. The model is rarely the limiting factor.
Where this starts
The demonstration always works
A retrieval system built in an afternoon answers beautifully across the ten documents it was built with. At ten thousand documents, spanning several years, containing three superseded versions of the same policy and a great deal of material nobody has ever curated, it begins returning answers that are fluent, plausible and wrong. Nothing about the model changed. What changed is that retrieval stopped finding the right passage.
What we usually find
- Answers drawn confidently from a superseded version of a document
- A fluent answer where the correct response was that the material does not say
- No citation, so nothing can be checked without redoing the work by hand
- Content sitting in formats the ingestion quietly skipped
- Chunking that split a table, a clause or a procedure down the middle
- No measurement, so quality is whatever the last person to try it recalls
Our position
Almost every disappointing retrieval system is a retrieval problem wearing a model problem as a disguise.
What the work covers
Ingestion, retrieval, grounding
The work is mostly upstream of the model. That is unglamorous, and it is where the difference between a system people trust and one they abandon is actually decided.
Ingestion and normalisation
Getting the material in at all: documents, wikis, tickets, shared drives and scanned files, with layout, tables and structure preserved rather than flattened into an undifferentiated wall of text.
Chunking that respects structure
Splitting on the document’s own boundaries rather than at a fixed character count, so a clause, a table or a procedure survives as one retrievable unit instead of two useless halves.
Hybrid retrieval
Keyword and vector search together with reranking, because pure vector search reliably misses exact terms: a policy number, a part code, a case reference, the things people actually search for.
Freshness and supersession
Knowing which version is current and keeping the withdrawn one out of the answer. In a regulated environment this is the single most damaging failure mode available.
Grounding and citation
Every answer carrying its source, and an explicit route to replying that the material does not cover this, which is a far better answer than a fluent invention.
Retrieval evaluation
A question set with known correct passages, so a change to chunking, embedding or reranking can be judged on evidence rather than on whether it felt better in three tries.
How quality is kept
Measured retrieval, not assessed retrieval
Everything here exists so that a change can be shown to be an improvement rather than argued to be one.
Question set
Real questions paired with the passages that should be returned, maintained as an asset. It is the only way to compare two retrieval configurations honestly.
Recall
Whether the correct passage was retrieved at all, measured separately from whether the answer read well. These fail for different reasons and need different fixes.
Grounding
Automated checking that the answer is supported by the passages it cited, so an unsupported claim is caught in the pipeline rather than by a reader.
Refresh
Scheduled reingestion with supersession applied, so the corpus stays current without anybody maintaining it by hand and eventually forgetting.
Gaps
Questions the corpus could not answer, logged and reported. It is the most useful list your knowledge owners will receive all year.
What you are left holding
- An ingestion pipeline that handles the formats you actually hold
- Hybrid retrieval with reranking, tuned against your own questions
- Answers that cite their source and can say the material does not cover it
- An evaluation set that makes a retrieval change measurable
- A standing report of questions your material cannot answer
Tell us what people keep asking and cannot find
We will tell you whether your material can answer it at all, which is the question that decides the project.