The reliability layer that plugs into your existing RAG, measures every answer and makes it better every week.
How long do I have to get a refund for an online order?
« …you have 14 days to exercise your right of withdrawal… »
« …the refund is issued no later than 14 days after your request… »
Do you ship to Switzerland?
« …delivery to mainland France, Belgium and Luxembourg… »
Yes, we ship to Switzerland within 5 business days.?
Final answer: I don’t know. This information isn’t in your documents.
npm install @wyrlo/sdkimport { Wyrlo } from "@wyrlo/sdk";
const wyrlo = new Wyrlo({ apiKey: process.env.WYRLO_API_KEY });
const res = await wyrlo.guard({ question, passages, generate });pip install wyrloimport os
from wyrlo import Wyrlo
wyrlo = Wyrlo(api_key=os.environ["WYRLO_API_KEY"])
res = wyrlo.guard(question=question, passages=passages, generate=generate)Documents change, your users’ vocabulary drifts, retrieved passages lose relevance. Nothing breaks, and no alert fires.
The worst part isn’t the error, it’s the confident error. A wrong, well-phrased answer costs more than an honest “I don’t know”.
And every fix is manual: read traces, guess the cause, tweak a setting, hope nothing breaks.
“refund” is also searched as “withdrawal”.
Retrieved passages are reordered; only the most useful ones remain.
Instructions hidden in the question or the passages are blocked.
Each sentence of the answer is matched to the passage that supports it.
Certain, probable or uncertain: answer, or say “I don’t know”.
91.2%correct answers
+7.1 pts compared with week 37 (84.1%)
| Week | Correct answers | Version |
|---|---|---|
| S31 | 76.4% | — |
| S32 | 78.0% | — |
| S33 | 79.5% | — |
| S34 | 80.1% | v12 |
| S35 | 81.8% | — |
| S36 | 82.6% | v13 |
| S37 | 84.1% | v14 |
| S38 | 91.2% | v15 |
| Vocabulary | + “credit note” = “voucher” | 41 questions with no passage | +3.9 pts |
|---|---|---|---|
| Passages kept | 8 → 6 | diluted answers | +2.1 pts |
Your RAG
WYRLO, hosted in the EU (Paris)
| Free | Pro | Team | |
|---|---|---|---|
| Price | €0 per month | €49 per month, proposed [TO BE CONFIRMED] | €149 per month, proposed [TO BE CONFIRMED] |
| Availability | available | coming soon | coming soon |
| Questions analyzed per month | [TO BE CONFIRMED] | [TO BE CONFIRMED] | [TO BE CONFIRMED] |
| Extra credits | [TO BE CONFIRMED] | [TO BE CONFIRMED] | [TO BE CONFIRMED] |
| Nightly loop | [TO BE CONFIRMED] | [TO BE CONFIRMED] | [TO BE CONFIRMED] |
| Weekly report | [TO BE CONFIRMED] | [TO BE CONFIRMED] | [TO BE CONFIRMED] |
| Start for free |
No. WYRLO sits between your search and your answer. Your index, your model and your documents stay with you; you add a few lines of code.
No. WYRLO receives the question, the passage excerpts retrieved by your RAG and the answer. These traces are masked and kept 30 days by default, a retention period your organization can adjust. Your documents are never copied to our systems.
In the European Union: the database and trace processing are in Paris (eu-west-3 region). The list of subprocessors is in the data processing agreement [TO BE CONFIRMED].
Your RAG keeps answering. If WYRLO doesn’t respond in time, the library lets your RAG’s answer through without analyzing it [timeout TO BE CONFIRMED]. During the incident, you lose verification, not service.
[TO BE CONFIRMED: measured p50 and p95 latency]
It tests each settings change before applying it and keeps only those that improve the score without making a previously passing question fail [non-regression method TO BE CONFIRMED]. You can also approve each change yourself and roll back to a previous version.
[TO BE CONFIRMED: evaluation method]. The report always shows its base: the number of answers evaluated out of the number of questions analyzed.
WYRLO doesn’t modify your documents. It tells you the exact fix to apply, for example re-splitting a file by article, along with the affected questions.
A few lines of code, then your first measured answers today.
Four steps, then the loop runs on its own.
Once
1Install the library
A few lines between your search and your answer. Your RAG, your documents and your model stay the same.
On every question
2WYRLO measures every answer
Sources checked sentence by sentence, confidence level, decision: answer or say “I don’t know”.
Every night
3The loop adjusts the settings
It applies a change only if the score goes up without breaking anything.
Every week
4The report proves it
“91% correct answers, up from 84% last week”
Numeric example, demo data.