Askweft

Document clarification planning over extracted invoice and bank-event JSON. Askweft shows competing record groups, ranks field requests by exact expected elimination per declared cost, records typed answers and preserves unresolved cases in a local SQLite journal.

Python 3.11 or newer; no runtime dependencies.

Recorded example · Release packages

Run locally

git clone https://github.com/bglglzd/askweft.git
cd askweft
python -m venv .venv
.venv/bin/python -m pip install .
.venv/bin/askweft --db example.sqlite demo

On Windows use .venv\Scripts\python.exe and .venv\Scripts\askweft.exe. Open the printed numeric loopback URL. The synthetic example is seeded only when the database has no sessions. Ctrl+C stops the server. Use a fresh database path to start again.

Request Payment reference, answer PAY-42, then request Duplicate bank records and choose Yes. Inspect the surviving group and its two answer receipts. The hosted example lets you inspect six saved engine stages; local operation accepts your own inputs and answers.

Import your own Case v1 JSON through the local UI or askweft --db work.sqlite import case.json. Inspect the candidate partitions, explicitly request a field and provide its actual value or unknown. An answer outside all candidates stays unresolved. Restarting creates a new session from retained source bytes and preserves the old answers.

Scope

One invoice, up to 16 bank events and 16 explicit alternatives. USD/EUR/GBP; integer minor units; exact, case-sensitive comparisons. A group with multiple records describes duplicate representation of one payment, never a summed split payment.

One survivor is resolved within the supplied model. An unseen truth can mimic an included candidate. This is not document authentication, financial approval, OCR or a live bank feed. Question costs are declared units, not measured operator time or money.

The first synthetic comparison used 30 development, 10 held-out and 8 separate stress inputs. On the 10 held-out inputs at the standard cost setting, adaptive and fixed each resolved 7 correctly and left 3 unresolved; no wrong resolution occurred. Default held-out advantage over fixed was not established. Full effort, sensitivity, denominators and limits.

Case/request specification · Architecture · Operations and recovery · Evaluation replay · Related products · MIT license