“Level-A RAG for a consumer insurance explainer (no vector DB)”
Shipped an upgrade to CoverClear (insurance policy explainer under NexusDocs).
“We stopped truncating the front of insurance policies”
“Level-A RAG for a consumer insurance explainer (no vector DB)”
Problem we kept hitting:
A simple character truncate favors declarations and ignores exclusions that show up later. That’s exactly where people get surprised.
What we built (Level-A RAG, no vector DB yet):
- Redact common PII (SSN, email, phone, DOB, labeled IDs)
- Chunk on headings / natural breaks
- Retrieve by task — summary vs risks vs user question
- Cap context to the user’s tier limit and analyze with Grok
Same credits system as TermsGuard. Same UI. Better packing of the context window.
I’m not claiming “reads every page perfectly.” I’m claiming the model is more likely to see the sections that matter for coverage and exclusions.
Live: https://nexusdocs.ai/coverclear.html
Curious how others handle long-document RAG without jumping straight to embeddings + a DB.
Hands-on lesson from the CoverClear update: simple truncation is rough on insurance policies.
We used to read from page one until the character limit. Exclusions often sit later.