Shipped Level-A single-doc RAG on BillClear (medical bills) — section + redact, no vector DB
Build update: BillClear on NexusDocs.
Problem: A dense hospital statement or EOB often puts patient responsibility, adjustments, and “miscellaneous supplies” after the character budget if you only truncate from the top.
What we shipped (Level-A, one document at a time):
- Redact common PII before the model
- Chunk the bill
- Retrieve sections for summary/flags / questions
- Stay inside the user’s tier character limit
- Return plain-English summary + practical “ask billing” flags
No multi-doc vector store. No “we archive your medical file” pitch. Process for analysis: redact first, explain second.
Same credit pool as CoverClear and TermsGuard.
Still not medical or billing advice — a clearer first pass so people know what to ask.
https://nexusdocs.ai/billclear.html
The redaction-first approach is the part that stands out. For something handling medical bills, limiting the system to one document and removing common PII before analysis seems like a sensible way to make the first step more controlled.
Thanks — that’s exactly why we put redaction first. Medical bills almost always include account numbers, member IDs, and contact details. We didn’t want those patterns riding along into the model if we could strip the obvious ones up front. Keeping it to a single document also forced a simpler design: section the bill, redact common PII, retrieve what matters for the summary/flags, then explain. No vector store to maintain for v1, and a clearer trust story for anyone pasting something sensitive. This is also built into our other app, CoverClear, that analyses Insurance policies.
That makes the trust boundary much clearer. I’d be curious how people respond once they start using the redaction-first workflow on real bills.
Thanks — that’s the question we’re watching too.
We’re early. A small number of people have signed in; not many have run a real bill through the redaction-first flow yet, so I don’t have a clean “here’s how users react” sample.
What I can say from our own tests and the first real passes:
Redacting obvious PII first feels like the right default. Nobody has asked us to send the raw account number to the model.
The drop-off is still earlier: getting someone to upload a bill at all. Once they do, the summary/flags are what they talk about—not the redaction step, which is mostly invisible if it works.
We don’t have on-site reviews wired up yet. Feedback so far is off-platform (SaaSHub, a few comments, people we know).
So the trust boundary is clearer in the architecture than it is in the usage data. Next useful signal for us is whether someone comes back with a second bill without being nudged.
If you try a real statement on BillClear, I’d genuinely like to hear what felt safe vs what still felt like too much to paste.
That’s a useful update. I’d be interested in continuing the conversation beyond the thread — would you be open to sharing the best email to reach you on?