Specialised document system with strong traceability
- 200-500documents per case file
- 7agents, up from 3 in the PoC
The problem
A typical case file carries 200 to 500 documents, and complex ones run into the thousands. Reading, sorting, prioritising and synthesising them took weeks.
And the real difficulty is not retrieving content: it is separating signal from noise and backing every statement in the summary with its source.
The architecture
A multi-agent architecture that grew from 3 agents in PoC to 5 in pilot and 7 in the advanced version, with extra validation and citation reinforcement.
Document classification, key-document selection, structured synthesis and an explicit link between every statement and its evidence. It returns two artefacts: a structured index of the case file and a traceable executive summary.
Impact
- Case files of 200 to 500 documents routinely, and thousands in complex operations
- The architecture scaled from 3 agents in PoC to 7 in the advanced version
- Executive summary with statement → source traceability
- No statement without evidence behind it
Stack
- RAG
- Azure AI Search
- Python
- Document Classification
- Traceability