Shipped: a drop-in that makes LLM answers traceable to your sources (and drops the rest)
Background: I had a deterministic "quote-checking" engine running inside a reading app I shipped — it only lets the model quote, then checks each quote against the source text in code. I kept thinking every RAG app needs this, so I pulled it out into a gateway.
MaxModel is an OpenAI-compatible API with a verified mode. You pass your own sources; it returns the answer with verbatim citations + a list of claims it couldn't ground (dropped before they reach your users). The check is character-for-character against your text — not an LLM judging an LLM.
Live now: npm i maxmodel / pip install maxmodel · docs https://docs.maxmodel.com
What I'm trying to learn before going deeper:
- Is verifiable trust the wedge, or just a nice-to-have on top of the cheap gateway?
- Drop-in SDK vs. people wanting the whole backend (DB/auth/RAG) too?
- What's the "I'd pay for this" moment for you?
Happy to set up design-partner calls with anyone building support bots / internal copilots / anything that can't afford wrong answers.
hit the same thing building a citation app. worst bugs weren't made up, they were grounding on something real but adjacent and saying it with full confidence. character-level check is really the only thing that catches that. and yeah it's not a nice to have, once you've seen "confidently wrong" up close, plain citations stop feeling like enough.
One thing I'd be careful with is that the thing users value most isn't always the thing they're actually buying.
Verification may end up being extremely important once someone is using the product.
That doesn't automatically mean it's the reason they decide to adopt it in the first place.
Those can point you toward very different conclusions about the wedge.
That's a sharp distinction. Thanks, you've actually put a name to what I've been unclear about. You're right that "valued once you're in" and "the reason you adopt" can lead to very different products, and I've been mixing them up.
Let me ask you directly, since you clearly think about this. In your experience, does verification ever work as the main reason someone tries a product? For example, has anyone been burned by a confidently wrong answer in production and then started looking specifically for "prove it's grounded"? Or does that kind of pain only show up after adoption, so the real wedge has to be something more immediate, like price, drop-in compatibility, or latency?
I'm trying to figure out if "trust" is a vitamin or a painkiller in this case, and for which users.
That's actually the part I'd be hesitant to answer confidently in a thread.
What's interesting to me isn't whether a case can be made for verification as the wedge.
It's whether the current evidence has earned that conclusion over the competing explanations.
I've got a few thoughts on that, but it's probably more than I'd try to unpack properly here.
What's the best email to reach you on?
I appreciate your nuanced approach. Framing the discussion as "can a case be made" versus "has the evidence earned it" is effective, and I am genuinely interested in your perspective.
You can reach me at ynez@maxmodel.com whenever you have a moment. There is no rush.
Sent 👍