We built the thing that tells the AI "no."
FactQL sits between the model and your data. It checks every answer against the rows before you see it, and declines instead of guessing. Every answer comes with evidence: what it looked at, the rows it found, how it got there. 🧵
Don't build AI chat for your app. Paste this instead:
<script src="https://t.co/PCKt9I8Ubp"></script>
<factql-chat api-key="fl_live_..." end-user-id="user-42"></factql-chat>
Works in React, Next.js, Vue and plain HTML.
Someone asked our AI for every user's password.
3 prompts. 0 leaks.
The fix isn't a smarter refusal. It's limiting what the model can reach: FactQL only touches the tables you switch on.
What would you have typed?
You're not a developer. You're the team's human database.
Sales wants a number. Support wants a number. Finance wants theirs by 5.
Every "quick question" is a query you stop coding to write.
What if they could ask the data themselves and see the rows?
5 SQL mistakes that return the wrong number with no error:
1. Cancelled orders counted
2. A join that double-counts
3. The wrong "revenue" column
4. A table that stopped syncing
5. A time zone that shifts the month
Reproduce each in Postgres:
https://t.co/dpvozM2sq2
#SQL
Your agent's data tool comes back empty. It writes a confident number anyway.
New tutorial: a LangChain v1 agent whose data tool returns the query and rows behind each answer, or DECLINED, and the prompt line that stops the guess.
https://t.co/uTt81VwBe5
#LangChain#AIAgents
@divaagurlxw Point 4 is the one that bit us. Worse than an empty body: a tool that returns 200 saying "can't answer", and the agent smooths it over with an estimate.
Treat that as its own outcome: a prefix the agent can't miss, plus a prompt rule to say so instead of filling the gap.
@fullstacklover Write: no, not directly. Have it propose the change as a migration or PR and let a human merge it.
Read deserves care too: its own SELECT-only role, a statement_timeout, only the schemas it needs. The quieter risk with read isn't damage, it's a confident number nobody checks.
FactQL is on LinkedIn now.
First post: every "AI on your data" demo runs on 5 tables. Real databases have 400, three columns called revenue, and a soft-delete flag nobody documented.
That's where a plausible answer turns into a wrong number.
@kmeanskaran@googlecloud "Agents never guess from CSVs" should be the default for every data agent.
Curious what the Analyst does when the data simply can't answer the question. Does it say so, or still reach for the closest feature? That's where we see most confident wrong answers.
@rohanpaul_ai "Every table traces back to the code" is the part most analytics agents skip. People don't need to trust the model; they need to see where the number came from.
Same principle behind FactQL: each answer comes with what it looked at, the rows it found and how it got there.
@beamnxw "No verification step" is the line. Data questions are the worst case: the query runs, the number is wrong, nothing errors.
That's the layer FactQL fills for AI apps: it runs the query, checks the answer against the rows, and declines instead of guessing.
@SmallwoodAl_ Point 2 is the one most stacks skip. If an analyst can't see where a number came from, they re-check the whole note anyway.
It's the rule FactQL is built on: every answer shows what it looked at, the rows it found and how it got there, from Postgres, a CSV or a PDF.
@akshay_pachaar Ontology fixes what the columns mean. Two gaps remain: the question nobody certified yet, and data outside the warehouse, like a CSV export or a PDF.
FactQL covers both: Postgres, MySQL, DuckDB, CSV/XLSX and PDFs behind one API, and it declines instead of guessing.
@SuiBuilds And the five definitions usually differ by just enough to look plausible. A broken query gets fixed in a minute; a wrong number that looks right ends up in a board deck.
When they disagree on your team, whose definition wins: finance's or the dashboard's?
@_jaydeepkarale "Should cancelled orders count?" is the one that bites. The SQL runs, the chart shows a record month, nobody questions it.
We built FactQL for exactly that moment: it runs the query, checks the answer against the rows, and declines instead of guessing when it can't back it.
That green bar looked like a record month. $1.2M.
Then someone checked the actual orders. Some were cancelled.
Remove them and it's $0.9M.
Impressive isn't proof. FactQL checks the answer against the rows.
AI on Trial #002
Free tier, no card: https://t.co/gstupO6WZz