How we fixed long-document analysis with Level-A RAG

Updating CoverClear was a solid hands-on lesson in Level-A RAG.
Before: the app basically read the policy from the top and stopped at the character limit (about 5,000 on free, 50,000 on paid). On a long policy, that often meant declarations and limits were in scope, while important exclusions and duties further down never made it into the analysis.
Now: instead of only using the front of the document, CoverClear:

Redacts common PII (SSN, email, phone, DOB, labeled policy/claim IDs) before the model runs
Chunks the policy into section-sized pieces
Retrieves the parts that matter for the job—coverage summary, exclusions/gaps, or the user’s question
Sends those excerpts within the same tier character budget

So the limit is still there for cost and latency, but the budget is spent on relevant sections, not just “whatever was on page one.”
On privacy: if sensitive patterns are in the file, they’re stripped on the server before analysis. You may still see the original text in the upload/paste box—that stays on your side in the browser. Redaction happens when the document is sent for processing, not by rewriting the text area.