Corpus
Which documents are in, who owns them, and how a deleted file leaves the index.
Answers need a source.
Retrieval-augmented generation is how a model uses your material without pretending to have memorised it. The quality is in the index, the permissions, and the cases you rerun.
- Chunking is a design choice
- Permission travels with the chunk
- Evals use your questions
What retrieval has to get right.
A bigger context window is not a substitute for knowing which passage was allowed.
Query
The question the user asked is not always the query that should hit the index. Rewrite it on purpose, and test the rewrite.
Citation
Show the passage. If the interface hides it, you cannot review the answer.
Evaluation that can fail.
An eval is a test suite for answers. It needs known goods and known bads.
Grounding cases
Questions whose supporting passage you can point at. The answer must use it, not a neighbouring myth.
Refusal cases
Questions the corpus cannot answer. A confident invention is a failed test.
Permission cases
A user who should not see a document must not receive it through a paraphrase.
RAG, prompt, or fine-tune.
Pick the cheapest mechanism that can be updated when the facts change.
Prompt
Enough for format, tone, and a stable instruction. Not enough for a library of policy.
RAG
The default when the facts live in documents you will edit.
Fine-tuning
A later choice for style or a narrow classification, after you can already measure the task.
Bring ten questions you wish it would answer.
We can tell you whether you have a retrieval problem, a permission problem, or no corpus yet.