INSURANCE
Insurance Claims RAG Assistant
A retrieval assistant over policy and claims documentation, answering in plain language with the source passage attached.
- PROJECT
- Insurance Claims RAG Assistant
- INDUSTRY
- Insurance
- STACK
- Python · RAG · OpenAI · Anthropic · PostgreSQL · pgvector · LangChain
OVERVIEW
The project
Claims professionals were answering coverage questions by searching scattered files and remembering past cases. Allzone built a retrieval assistant over the organisation's own policy and claims documentation, so a question asked in plain language comes back with an answer tied to the specific policy language that supports it.
CHALLENGE
What made it difficult
In claims work an unsourced answer is not usable. A model that summarises coverage confidently but cannot show which clause it read leaves the person accountable for the decision no better off than before it answered.
The material is uneven too. Policy wordings, endorsements, handbooks and historical case files are written and structured differently, and the retrieval layer has to cope with all of them without a specialist reformatting each document first.
WHAT WE DID
The work delivered.
- 01
Document pipeline
Ingestion, parsing, chunking and embedding across policy and claims documentation, so the corpus is retrievable rather than merely stored.
- 02
Vector retrieval
Vector search on PostgreSQL with pgvector, retrieving the passages that actually support an answer to the question asked.
- 03
Grounded generation
Retrieval kept separate from generation, so the model reasons over retrieved evidence and the answer carries citations back to it.
- 04
Monitoring and audit
Observability over retrieval and response quality, with audit controls covering what the assistant answered and which source it drew on.
ARCHITECTURE
How it fits together
- A reusable pipeline from source documents through embeddings to vector search
- Retrieval separated from generation, which is what makes the evidence behind an answer traceable
- PostgreSQL with pgvector as the vector store, alongside operational data rather than in a separate system
- Observability over retrieval quality and answer grounding for production monitoring
TECHNOLOGY
The stack
- Python
- RAG
- OpenAI
- Anthropic
- PostgreSQL
- pgvector
- LangChain
OUTCOME
What exists now.
A claims professional checks an answer against the clause it came from, rather than taking a summary on trust
Institutional knowledge is available to everyone handling claims, not only to the people who have been there longest
Retrieval behaviour and answer quality are monitored in production, so a drop in grounding is visible instead of silent
NEXT STEP
Tell us what you're building.
Bring us the problem with its real constraints attached. We will tell you what we would build, and what we would not.