# Useful retrieval is a system, not a vector database.

> Help an AI answer from relevant, current sources rather than memory alone. Check what it finds, whether the source supports the answer and how quickly it responds.

Canonical URL: https://marmation.com/expertise/rag-and-retrieval-engineering/

## The decision

Will a support answer include the exception that changes what the customer should do?

Check the source passages before judging the answer. Search must keep rules with their qualifications and use the right document version; fluent writing cannot repair missing evidence.

## A support answer loses the exception that changes the advice

A technical support organization has manuals with tables, versions, diagrams and exceptions. A naive text chunker separates the qualification from the rule and causes a confident wrong answer.

## Missing source meaning cannot be repaired by fluent writing

Weak parsing, chunk boundaries, missed metadata and ambiguous source authority can hurt quality before the model ever reads the material. Longer contexts are not a substitute for precise evidence.

## Test how each source passage reaches the answer

Start with support questions and the passages needed to answer them correctly. Test document extraction, how text is divided into searchable sections (chunking), keyword and meaning-based search, and how candidate passages are reordered for relevance (reranking). Preserve tables, exceptions, citations and document versions. Measure the missing or incorrect evidence separately from the final answer, and inspect failures by hand before tuning the search.

## The extraction, search and citation work the task may need

- Parsing and provenance pipeline for target document types.
- Hybrid retrieval/reranking design with filters and source metadata.
- Answer grounding and clear citation interface.
- Quality, freshness, permissions and latency evaluation.


## Use managed retrieval until a specific failure justifies more

Built-in managed RAG may be enough for simple sources; bespoke retrieval is more useful for specialized document structures, fine-grained filtering or domain-specific ranking.

## Check exact values, versions, inaccessible sources and missing answers

Test exact-value lookup, cross-document synthesis, missing answers, version drift, inaccessible documents and corrupted ingestion. Record source passage quality separately from answer fluency.

## Write the questions and expected evidence before tuning search

Build an evaluation set before optimizing search. Without known queries and expected evidence, better retrieval cannot be distinguished from a more persuasive answer.

## Related reading

- [Retrieval cannot be allowed to widen a user's permissions.](https://marmation.com/expertise/permission-aware-rag/)
- [Company knowledge should be useful to the people allowed to use it.](https://marmation.com/solutions/enterprise-knowledge-systems/)
- [RAG is not a custom-trained model. That distinction can save a project.](https://marmation.com/review/rag-is-not-a-trained-model/)