Answering policy and expense questions on-device
An HR or finance team can put up to 40 documents — employee handbook, expense policy, vendor contracts, invoices exported to CSV — into one browser-local library and question them together. Nothing is uploaded. This page covers the supported formats, the question patterns that work, and the governance questions a local tool does not answer.
Last updated:
Most of the value sits in three recurring questions: what does the policy say, which contract clause applies, and does this invoice match the agreed price. The first two are retrieval problems. The third needs the numbers present as text.
Formats, including one gap worth knowing
PDF, DOCX, TXT, Markdown, CSV, HTML and JSON are supported directly. Spreadsheets are not parsed as .xlsx, so an expense log has to be exported to CSV first — a single click in most suites, and a step people forget until a library comes back with nothing useful from their biggest file.
| Source | How it goes in | Why that matters |
|---|---|---|
| Employee handbook .docx | Import as is | Heading levels are preserved, so Section 4.2 keeps its parent |
| Expense policy .pdf | Import as is | Page numbers come through, so a citation reads “page 3” |
| Invoice bundle .xlsx | Export each sheet to CSV | Spreadsheets are not parsed; a CSV keeps row boundaries intact |
| Vendor master held online | Download as CSV | The same export works offline once cached |
| Scan-only receipts .pdf | OCR first | Image-only pages yield no text at all |
| Mixed audit working papers | Merge into one PDF where possible | Fewer files keeps you under the 40-file cap |
Three question patterns, with examples
- Policy lookup — “What receipt does a taxi claim of 800 CNY require, and which section says so?” The answer returns the section with its heading path and a page number.
- Clause location — “Which vendor contracts have a 30-day termination for convenience clause?” Both phrases are exact tokens, so keyword ranking surfaces them and the answer quotes each clause with a file name.
- Reconciliation — “Which invoices here bill above the unit price stated in the contract?” Needs the CSV and the contract loaded in the same library. Retrieval returns both figures; the subtraction stays with you.
Two hours of audit prep, step by step
Collect the paper equivalents in one folder
Handbook, expense policy, the three vendor contracts under review, and one quarter of invoices exported to CSV. Ten files, comfortably under 200 MB.
Import and watch the index build
Parsing starts with magic-byte sniffing rather than the extension, then format-specific extraction, then chunking at about 700 characters with 15% overlap. The progress panel names each file and reports the ones it cannot read.
Ask in the wording employees actually use
Staff ask “can I claim a taxi home after 9pm?” while the policy says “late-night transport”. Ask it both ways: hybrid retrieval exists for exactly this vocabulary mismatch.
Verify two answers against the source
Click each citation and confirm the section number and the policy revision date. Nothing checks whether the version you loaded is the current one.
Handle the spreadsheet separately
Export to CSV, then import. Confirm the delimiter was detected; a misread delimiter turns a six-column file into one column of noise.
Save the outcome, not the index
Export the question and answer list with citations to Markdown for the working papers. The index itself is disposable and disappears when site data is cleared; the export survives.
What a local tool does not solve
- Data classification. Which categories of employee data may be processed at all is a policy decision, and it is not settled by where the computation happens.
- Access control. There are no accounts and no permission layer: whoever has the unlocked device and that browser profile has the library.
- Device loss. Clearing site data deletes the index, which is good for disposal and bad if the only copy of your notes lived there.
- Shared machines. A library in a shared browser profile is readable by the next person who opens it. Use a separate OS account, or clear site data when you finish.
- Retention rules. Nothing is logged, so there is no automatic evidence that a check was performed. Take your own notes.
When this is not the right approach
- You need a central, permissioned repository with an audit log. That is a different class of system, deliberately.
- The corpus exceeds 40 files or 200 MB. A full invoice archive does not fit: split by quarter, or use the accounting system’s own search.
- Scanned receipts are the core of the work and no OCR step exists in your process.
- Two people must work from the same live index at once.
- You need numbers extracted and totalled automatically. Retrieval returns passages; arithmetic on them is yours, and a wrong total from a language model is worse than no total.
- The answers would be filed as controls evidence without a human reading the cited passage.
Verifying the privacy claim yourself
- Open the tool beside DevTools → Network and clear the log.
- Add a policy PDF. Parsing, chunking and indexing produce no request at all.
- Ask three questions in the default retrieval tier. Still no request.
- Only when you opt into the generation tier does one model download appear — 400 MB to 1.0 GB behind a click-to-confirm dialog — then switch DevTools to Offline and keep working.
- Clear site data and confirm the library is empty.
Can it read our expense spreadsheet directly?
Not as .xlsx. Export the sheet to CSV and import that; the parser supports CSV and JSON, and a CSV keeps the row boundaries that make “sum the amounts above 500” style questions possible.
Where is the index stored, and who can read it?
In IndexedDB in the browser profile you used. Anyone with that device and that profile can open it, so on a shared machine use a separate OS account or clear site data when you finish.
Does using this make us GDPR or PIPL compliant?
No tool makes a process compliant. Keeping processing local changes which transfers happen; the lawful basis, retention and access decisions remain yours.
How many documents fit?
40 files, 25 MB each, 200 MB total and 20,000 chunks per library. A policy set fits easily; an invoice archive needs splitting by quarter.
Can two colleagues query the same library?
No. The index is per browser profile with no server behind it. Share exported answers with citations, or have each person import their own copies.