Retrieval quality is an engineering result, not a model setting.
The retrieval layer underneath every grounded AI system: chunking, indexing, ranking and evaluation, built so the right passage comes back before anything is generated.
Where retrieval-augmented systems fail
Teams change models when the answers are poor. The answers are usually poor because retrieval returned the wrong passage, and no model recovers from being handed the wrong evidence. Retrieval is the part that decides quality, and it is the part that is most often left at defaults.
- The assistant answers well on some documents and badly on others, with no obvious pattern.
- Nobody can say whether a bad answer was a retrieval failure or a generation failure.
- Documents were indexed once and the index has drifted from reality since.
- Retrieval quality has never been measured, only observed anecdotally.
What we build
A retrieval system with an evaluation harness attached, so changes to chunking, embedding or ranking can be judged on evidence rather than on impression.
- Document processing that preserves structure rather than splitting on character counts
- Chunking and embedding strategy chosen against your content, not a default
- Hybrid retrieval combining semantic and keyword matching where it improves recall
- Re-ranking to put the genuinely relevant passage first
- An evaluation set and harness that scores retrieval separately from generation
- Incremental refresh so the index tracks the source as it changes
How it runs
Every stage is measurable, which is what makes the system improvable.
- 01Build the evaluation set
Real questions with known correct source passages. Without this, every subsequent change is guesswork.
- 02Process the corpus
Documents are parsed with their structure intact — headings, tables and sections carry meaning that flat text loses.
- 03Index and rank
Embedding and keyword indexes built together, with a re-ranking pass over the candidate set.
- 04Score retrieval alone
Did the right passage come back, in the top results? That question is answered before any model is involved.
- 05Refresh continuously
Source changes propagate to the index on a known cadence, so answers reflect the current state rather than the launch state.
What changes once it is running
What separating retrieval from generation gives you.
Failures become diagnosable
A bad answer is traceable to retrieval or to generation, so the fix targets the actual cause.
Improvements are provable
A change to chunking or ranking is scored against the evaluation set rather than argued about.
The index stays current
Refresh is engineered, so the system does not slowly drift away from the documents it claims to know.
Model choice becomes cheap
With retrieval solid and measured, swapping the generation model is a low-risk decision rather than a rebuild.
How an engagement is shaped
Evaluation first. It is the only way to know whether anything after it worked.
Corpus and evaluation
Two to three weeks assembling the evaluation set and establishing baseline retrieval quality on your content.
Engineer and tune
Processing, indexing and ranking built and iterated against the harness until quality plateaus.
Operate
Refresh pipelines, drift monitoring and periodic re-scoring as the corpus grows.
Common questions
The things buyers ask before they commit. If yours is not here, it is a good first question for the assessment.
- Do we need a vector database?
- Often, but not always, and not necessarily a dedicated one. Several databases you may already run support vector search adequately at your volume. We size this against your corpus rather than defaulting to a new product.
- How large does the corpus need to be?
- There is no minimum. Small corpora can be handled with simpler retrieval and often should be — introducing embeddings for two hundred documents adds cost without adding quality.
- Can this sit under an assistant we already have?
- Yes. Retrieval is a separable layer. Replacing a weak retrieval layer under an existing assistant is a common and comparatively contained piece of work.
Bring twenty questions and the documents that answer them.
That is enough to measure whether your retrieval is the problem.