Every connected system, in its native protocol. Filter by the group it's assigned to.

Upload a document once and it becomes part of a group's permanent knowledge base — every future question can draw on it. Different from a chat attachment, which only applies to that one message.

Turns a folder of PDFs into live, queryable data — PDF → AI Extract → Ground → JOIN → Answer. Backed by this project's own PdfFieldExtractor (OCR fallback via PdfOcrEngine when a page has no text layer) — no separate ETL pipeline, no pre-processing step.

SELECT x.po_number, x.amount_due AS pdf_amount, x.grounded, po.total_amount AS sap_amount
FROM PDF_EXTRACT('s3://po-documents/', schema => 'po_number,vendor,amount_due') x
JOIN sap_orders.purchase_orders po ON po.po_number = x.po_number
WHERE x.amount_due <> po.total_amount

A group is the access + LLM boundary a business user actually sees — one place to ask, not a list of systems to know about.

Which group(s) each user can ask questions within. A user sees only the sources their group grants — nothing else.

Reusable, admin-authored instructions for a recurring analysis workflow. Business users invoke one by typing /skill-name in Ask, or ThinkingSense loads it automatically when a question matches its trigger.

Reusable question phrasings a business user picks and fills in themselves — lighter-weight than a Skill (no matching logic, no instructions, just a starting shape for a common question). Values are never auto-filled; the user always fills in the {placeholder}s.

Knowledge surfaced from real conversations, not written from scratch — a business question's answer already states a fact worth keeping; this is where an admin reviews it before it becomes trusted, reusable knowledge.

Turns a Slack channel into a shared place for data questions — link a channel to a group's data area, then anyone in that channel can ask ThinkingSense by mentioning it, with the answer posted as a thread so the main channel stays readable.

Every real natural-language question asked of ThinkingSense, from either entry point — the built-in Ask app or Claude via MCP — lands in the same log. This is the raw signal Ontology review and Rollup suggestions are both scanned from.

Curated NL ↔ SQL pairs a business question can always trust — the admin-facing counterpart to this project's own QuestionBankGenerator/QuestionBankRetriever. Marking a pair reviewed means the exact wording and its SQL are vetted, not just plausible.

Every 👎 tagged in Ask lands here — the aggregated view of what's getting flagged, separate from any single answer's feedback thread.

A composite score per group — illustrative weighting, but every input is a real, live count from this console (ontology backlog, open negative feedback, rollup coverage, skills defined), not a fabricated number.

Relationships inferred from real query history, value overlap, and business-user feedback — nothing gets used by NL2SQL until reviewed here.

Pre-aggregated tables that transparently speed up matching queries — suggested automatically from real repeated NL2SQL history across every group.

Real federated-query plan history — measured per-backend timing, not just the planner's own pre-execution estimate.

Manual SQL exploration against any registered backend, independent of NL2SQL — for validating joins, checking row counts, and debugging what a generated query would actually return.

An isolated workspace for drafting, refining, and validating a query before it becomes a trusted, reusable answer — separate from live chat, so iterating here never clutters a real conversation.

Prompt sets with expected SQL, run against NL2SQL to measure accuracy over time — the same shape as this repo's own Nl2SqlEvalTest suite, exposed here for ad hoc admin runs.

A digital worker configured once and left running — retrieves data, runs analysis, and delivers results on a schedule instead of waiting to be asked.

Per-group display, scheduling, and chat-behavior settings.

Embed ThinkingSense in your own product, and manage programmatic access.

<iframe
  src="https://app.thinkingsense.ai/embed/chat?group=retail-operations&token=SESSION_JWT"
  style="width:100%;height:640px;border:0">
</iframe>

import { ThinkingSenseChat } from '@thinkingsense/react'

<ThinkingSenseChat
  group="retail-operations"
  getToken={() => fetch('/api/thinkingsense-token').then(r => r.json())}
/>

A user who's stuck can flag an admin directly from Ask — separate from 👎 feedback on a specific answer, this is "I need a person," not "this answer was wrong."

ThinkingSense is multi-tenant — this admin console is scoped to whichever organization is active, switchable from the topbar.

Same one place to ask, whichever group you're in — try all three teams.

Signed in as Sam Chen · Finance Ops — group: Business Process Investigation

This demo answers the pre-loaded scenarios above — try one, or type close to it. Type / for skills, Tab to accept a suggestion, or press ⌘K / Ctrl K anywhere to jump around.

Answers you've pinned from Ask, laid out as a living page. Viewed live in the app, every widget re-runs against the same governed pipeline — but a shared link serves a pre-computed snapshot instead, so opening it never re-queries your sources.