RAG
RAG development so answers come from your material
Retrieval-augmented generation lets a model answer from your policies, manuals, tickets, or listings instead of from memory. We build the ingestion, the permission filter, the retrieval, and the answer format, then test it on questions your team actually asks.
- Ingestion
- Permission-aware retrieval
- Answer layer
- Maintenance
Why most RAG projects disappoint
They dump every file into a vector index and hope. Permissions are ignored, scanned PDFs are unreadable, and two contradictory policy versions are both “relevant”. We start by deciding which collections are in scope, how a document is updated, and who may see each one.
What the system returns
A useful RAG answer cites the passage it used and says when it did not find enough. It does not blend three policies into a fourth. For staff tools, the citation is the feature. For customers, the wording is tighter and the refusal is polite.
Answers from your documents
Retrieval includes
- The collections a role may search
- A citation the user can open
- A refusal when the source is missing
Not implied
- The whole company drive on day one
- Answers that ignore permissions
- A guarantee the model is never wrong
What changes a RAG system
The corpus
Current procedures are in. Old drafts and personal drives are a decision, not a default.
Permissions
Search has to follow the same access as the files. A shared index that leaks is a failed release.
Freshness
Say how a replaced policy leaves the index, and who notices.
What a retrieval brief needs
- 01
The collections
Which libraries a role may search.
- 02
The permission
The same access rules as the files.
- 03
A good answer
Two examples you would accept, with the file they should cite.
- 04
A bad answer
A question it should refuse.
Marks retrieval is doing the job
- 01
A source is cited
An answer points at the file or passage it used.
- 02
Missing says so
If the document is not in the index, the assistant says it does not know.
- 03
The index rebuilds
A new or corrected file can be added without rebuilding the product.
- 04
Access follows files
A person only retrieves documents they are already allowed to open.
What we build
Ingestion
Parsing, chunking, and metadata for the formats you really have, including messy office files.
Permission-aware retrieval
A person only retrieves what they could already open in the source system.
Answer layer
Prompts, citations, and a refusal path when retrieval is weak.
Maintenance
A way to add, replace, and retire documents without rebuilding the product.
How an engagement runs
- 01
Choose the corpus
One collection with an owner, not the entire company drive.
- 02
Test retrieval before prose
If the right passage is not retrieved, a better prompt will not save it.
- 03
Open it to a pilot group
Collect the questions that failed and fix the corpus or the index.
Related reading
What is RAG
A plain explanation of retrieval-augmented generation.
Knowledge assistant
Staff answers cited from policies they are allowed to read.
AI assistants
In-product help for staff, grounded in the systems they already use.
Document management
Versions and permissions, so the current file is obvious.Questions we hear
What is RAG in plain language?
The model is given relevant excerpts from your material at question time, and asked to answer from those excerpts. Our guide What is RAG explains the pattern without the sales language.
Can this respect department permissions?
Yes, if the source system can tell us who may read a document. If permissions live only in people’s heads, the index will leak or we will refuse to launch it that way.
Does RAG need a vector database?
Often, alongside ordinary filters on title, date, and department. Vectors are not a strategy by themselves. Keyword and metadata still matter.