Ask a question in plain language — “cyber cases in Madhapur in 2024” — and an AI agent queries the crime database, builds the report, checks it, and returns a download. The whole flow is: question → checks → agent thinks → queries the database → builds the Excel/Word/PDF → quality checks → saves and returns.
Asking a question
You write the ask in natural language and choose the output format — xlsx (Excel), docx (Word), or pdf — and a mode:
- Agentic — the agent runs on its own and the file is saved to your library automatically.
- Regular — the file comes back inline so you can keep editing before saving.
Gatekeeping
Only signed-in users get through, and a daily quota caps how many reports each user can generate — keeping usage controlled and accountable.
The thinking loop
The “agent” reasons step by step, but within firm limits so it never runs away or returns nothing:
- Step budget — a cap on how many thinking steps it may take (40).
- Time budget — a cap on how long it can run before it must wrap up.
- Final stretch — when steps or time are nearly gone, it stops researching and just builds the file, so you never get an empty result.
- Tool calls — it doesn’t answer from memory; it calls tools to look things up and to build files.
Looking up the data
The agent works against the Cyberabad FIR records (~250,000 cases). It peeks at the schema (list tables, describe table, sample rows), then runs the actual SQL to pull the matching cases, and can use ready-made breakdowns by crime type, station, and more.
Building the deliverable
From the rows it gathers, it can create charts, an Excel sheet (including straight from a query, with no row limit), a Word document, or a PDF — the finished file is the artifact, with a small preview of rows shown in the chat.
Quality checks before handing it over
A second AI verifier reviews the answer before delivery:
- Citation check — confirms the case numbers it cites are real, not invented.
- Location check — confirms it actually filtered to the right area or station.
- Degeneration guard — catches junk output like “2 ? ? ?” instead of a real number and makes it redo that part.
- Data integrity warnings — surfaces any flags the checks raise.
Saving and delivering
Agentic files are stored in your library automatically with a clean, human-friendly filename and searchable keywords pulled from the question. You get a download link, and a step trace records every step and how long it took.
Frequently asked questions
How do I ask for a report?
In plain language — for example, “cyber cases in Madhapur in 2024” — and you pick the output format (Excel, Word, or PDF) and whether it runs agentically or returns inline.
How do I know the report is accurate?
A verifier runs citation checks (real case numbers), a location check (correct area/station), and a degeneration guard (no junk values) before the file is handed over.
What data does it use?
The Cyberabad FIR records — around 250,000 cases — queried with SQL, plus ready-made breakdowns by crime type and station.
Will it ever return an empty result?
No. When the step or time budget is nearly spent, the final-stretch rule forces it to stop researching and build the file so you always get a deliverable.