Projects / rag-knowledge-assistant
Enterprise knowledge assistant (RAG)
A chat assistant that answers from internal document libraries and optional web search, cites its sources and can return a Word report.
- Period
- 2025–2026
- Role
- GenAI engineer
- Context
- Large industrial group
- Cloud
- Azure
[ Architecture ]
How it works
Run a scenario to follow a request through the system. Select a component to read its role and its flows.
an employee
What are the steps of the inspection procedure?
A question about an internal procedure, asked in the company chat.
- Press Ask about a procedure to watch the request go through the system.
- 01EmployeeAsks the question in the chat
- 02Company chatMessage relayed to the assistant
- 03RAG assistantConversation loaded · plan - search internal documents
- 04Hybrid searchKeyword and vector search · top passages returned
- 05RAG assistantPassages ranked and kept
- 06LLMAnswer written from the passages · sources cited
- ✓The employee gets an answer in the chat with its sources listed, internal documents first.
- 01EmployeeAsks for a report in the same conversation
- 02Company chatMessage relayed to the assistant
- 03RAG assistantEarlier answer and sources reloaded · Word document built
- 04Company chatDocument delivered in the chat
- ✓The report comes back in the same conversation, built from the earlier answer and its sources.
Problem
Employees needed a conversational way to query knowledge scattered across heterogeneous internal document libraries. The assistant also had to offer optional web search, produce a document as output, run inside the existing Azure environment and show its sources.
Data
Internal documents of several kinds, public web pages when the web option is used, and the conversation history. Access follows the existing rights on the libraries.
Approach
- Keyword and vector search run in the same PostgreSQL database, so exact terms and paraphrases both match.
- Internal documents come first, web search second, and the answer cites where each part came from.
- The steps are a LangGraph graph, with memory per conversation and a Word report generator for follow-ups.
Evaluation
No formal evaluation set was recorded for this project. The design makes answers checkable instead: each answer cites the passages it relies on.
Result
The assistant runs inside the company chat app, and its answers cite their sources.
My role
GenAI engineer on the whole chain: use-case framing, RAG architecture, backend, retrieval, orchestration, chat integration, Azure deployment decisions and user onboarding.