Make data and AI useful
Get answers from your own documents, with the sources
An assistant that searches the documents your team is allowed to see and answers with references, so people spend less time searching through folders and old threads.
In short
A knowledge assistant uses retrieval-augmented generation (RAG). It searches only the documents a person is authorized to see, gives the model that context, and returns an answer that cites its sources where retrieval supports them. When the evidence is thin, it says so and points to a person.
Last updated
01 The problem
What problem does this solve?
Policies, manuals, runbooks and past answers live in different tools. Finding the right one takes time, so people ask the colleague who knows instead, and that colleague answers the same questions again and again.
A general chatbot cannot see your documents, and pasting them in raises access and privacy questions. A useful assistant has to respect who is allowed to see what, and say when it does not know.
02 What we build
What do we build?
- Document ingestion with parsing, chunking and metadata enrichment
- Retrieval that searches only what each user is authorized to see
- Hybrid search and reranking, tuned on your real questions
- Answers with source references and an escalation path
- An evaluation set of representative business questions
- Monitoring of answer quality, latency and cost per answer
03 Architecture
What does a typical architecture look like?
04 How it goes
How does the work usually go?
-
Collect the real questions
We gather the questions your team actually asks and agree what a good answer looks like. These become the test set.
-
Prepare the sources
We inventory the documents, confirm who may see what, and set up ingestion that keeps the index current.
-
Build and test retrieval first
Retrieval quality decides answer quality, so we measure it against the test set before tuning prompts or models.
-
Pilot with agreed criteria
A time-boxed pilot with success criteria, known limitations and a clear decision to proceed, revise or stop.
-
Launch and keep measuring
We release with monitoring of quality, latency and cost, and review the results as documents and questions change.
05 Capabilities and stack
Which capabilities and technologies are involved?
Capabilities
- LangGraph
- LlamaIndex
- PostgreSQL with pgvector
- OpenSearch
- Amazon Bedrock
- Google Vertex AI
- FastAPI
The final stack follows your requirements and what you already run. How we choose technology
06 A typical situation
What does this usually look like?
A services firm with years of proposals, policies and project notes depends on the few people who know where things are.
A pilot on one team's documents, tested against their real questions, shows whether an assistant helps.
07 Questions
Questions people ask
Retrieval is filtered by the signed-in user's permissions before the model sees any content, and we test for unauthorized disclosure as part of the evaluation before launch.
Not by default. A hosted foundation model with good retrieval is the starting point, and we consider fine-tuning only where testing shows retrieval alone cannot reach the quality you need.
The assistant should say the evidence is insufficient and point to a person or team instead of guessing. We include these cases in the test set.
09 Terms
Which terms come up here?
If this sounds like you: Answer repeat questions from your own help content