PromptQL: a context wiki instead of a semantic layer

Definition

PromptQL is Hasura’s product for asking questions over company data. It started as “accurate text-to-SQL through a semantic layer” (an explicit query plan over Hasura’s supergraph metadata) and by late 2026 had rebuilt itself around a different claim: the thing that makes an agent accurate is not a schema catalog but a wiki of operational context that grows in the flow of work. Their page says it outright: semantic layers and knowledge graphs “don’t scale”; definitions, exceptions and tribal rules do.

Rule

Seven mechanics, as they present them:

1. Context wiki, not a graph. A page per metric or process: formula, policy rule, categories, regional targets, exceptions. Tags and a Talk tab. The agent reads these pages before it reads the data.
2. “PromptQL wants to learn.” A user states a rule in a thread. The agent drafts the wiki edit, shows it, and the user reviews, edits, then clicks “Add to wiki” or tags someone who knows. Nothing lands without that click.
3. AI flags gaps. When context is missing, the agent says so and suggests the page to write, instead of answering confidently from nothing.
4. Versioned, attributed, reversible. Every edit has an author, a revision history and a rollback. They sell this as governance; it is also what makes the wiki trustworthy enough to act on.
5. Repeated work becomes a program. A workflow that is asked twice is saved and re-run, not re-planned. Also their cost lever: no regenerating plans.
6. Permissions per request. The bot has no access of its own; each action borrows the permissions of whoever asked. A shared bot is not a side door.
7. Feedback on the answer. Confidence shown; Approve / Inaccurate buttons. Plus multiplayer threads where people @-mention each other and the bot.

Exceptions

What did not transfer, and why:

• Multiplayer threads and per-request permissions: one person, one wiki, one git identity. There is nothing to arbitrate.
• Saved programs: already exist here as skills and make targets. The missing piece was only the trigger, “asked three times → write it”, which the gaps log gives.
• Confidence on answers: a number without a calibration is decoration. Logging the gap says the same thing with evidence attached.
• Their original query-plan idea (plan as explicit Python over semantic metadata, visible before execution) is still good and still absent from their marketing. It is the same insight as schema-guided-reasoning: make the plan a typed object, not a hidden chain of thought.

Numbers

• Pricing, 2026-10-07: Free playground; Team $40 per user per month; Enterprise from $1,000 per month, single-tenant; Custom with BYOC and forward-deployed engineers.
• Landing page, docs overview and “why it works” page carry zero architecture detail: no plan format, no benchmark, no accuracy figure. Everything above is mechanics they show in demos, not measured claims.

Learned from

Fetched promptql.io, /accurate-ai, /why-promptql-works and /docs on 2026-10-07 while asking what our LLM wiki was missing. Taken: mechanics 2, 3 and 4, implemented as files, grep and git in wiki-feedback-loop. The shape rhymes with agent-mistake-fix-harness: a correction is a harness defect, and here the harness is the wiki.