Skip to main content
Not yet built. The catalog and profile pages Ask ATX reads from are live, but the query box, the retrieval pipeline and the grounded answer are still ahead. This page describes the design.
Ask ATX is the front door — a natural-language query box sitting on top of the public directory.
“LLM providers that are SOC 2 and ISO 42001 certified and don’t train on customer data”
A reviewer types the question they actually have, and gets a short summary plus a ranked vendor shortlist where every factual claim links to the specific field it came from and the date that field was last verified.

​
The pipeline

1

Query understanding

The natural-language question is split into structured filters — category, certifications, compliance posture, data flows, deployment model — plus the free-text intent that no filter captures.
2

Hybrid retrieval

A structured filter runs over the canonical catalog, a semantic retrieval runs over profile text via pgvector embeddings, and the two result sets are combined and reranked. Structured alone misses paraphrase; semantic alone misses hard constraints like a named certification.
3

Grounded answer

A model composes the summary and the shortlist from retrieved fields only. Each claim cites its source-cited field and its verified date; answers are framed as grounded opinion, never as an unattributed verdict.

​
It sits inside the firewall

This is the part that is easy to get wrong. A conversational recommender that ranks on anything a vendor can buy is a pay-to-play engine wearing a chat interface — and it carries exactly the same defamation and independence exposure as a bought rating, with less of the scrutiny. So the same rules that govern a published rating govern the answer box:
  • Retrieval and ranking draw only on independent trust data — verified and crowd provenance, and published ratings.
  • Vendor-claimed commercial fields and any vendor spend are never ranking inputs.
  • Every factual claim cites its field and its last_verified_at.
  • Fields past their re-verification window are flagged in the answer, not silently presented as current.
This is enforced by isolation tests asserting that no commercial datum can enter the retrieval or ranking path — a test that proves the boundary, rather than a comment claiming it. See Independence.

​
Quality is measured, not asserted

A grounded-answer system fails quietly: it keeps producing fluent, plausible answers as its retrieval degrades. The design guards against that with a fixed evaluation set of canonical due-diligence queries and their expected shortlists, run in CI — so a regression in retrieval shows up as a failing build rather than as a slightly worse answer nobody notices.
The first version answers a single question at a time. Multi-turn refinement — follow-ups and session memory — is a planned follow-on, not part of the initial capability.

​
Queries as signal

Questions asked of Ask ATX are captured in anonymized form as buyer-intent signal: what the market is evaluating, against which constraints, and when. It is a genuinely valuable data asset, and it is covered by a privacy notice that says plainly what is retained.

Trust profiles

The fields every answer is grounded in, and how their verified dates are maintained.

Independence

Why the ranking path and the commercial path are separate services.