askrobots / dbbasic-object-server
Public
Runs live, versioned Python application objects on one VM, with a suite of apps: projects, notes, tasks, files and more.
danq9
signup: the person chooses the password and name on the confirm page
83c7aa6
2 weeks, 5 days ago
History
scans.json · 5.2 KB · 58 lines
| 1 | { |
| 2 | "name": "scans", |
| 3 | "title": "Scans", |
| 4 | "version": 1, |
| 5 | "description": "A photographed or emailed document on its way into the books. The record is EVIDENCE and is never edited (docs/logic-decisions.md #8): the image is what an auditor actually wants, the OCR text is what a machine read, and the extraction is what a machine GUESSED. None of it posts anything. The write that matters happens when a human confirms, which creates a draft expense (or bill, or bank import) stamped generated_from scans/{id} and attributed to them -- so a weak extractor costs somebody thirty seconds of typing rather than putting a wrong number in the ledger. That boundary is why this can ship with no AI at all: the free path reads totals with regex over OCR text, an AI vision pass simply guesses better, and confirmation cannot tell them apart. content_sha256 makes re-snapping the same receipt land on the same scan, because the second photo of a crumpled receipt is the normal case, not the exception.", |
| 6 | "fields": [ |
| 7 | {"name": "id"}, |
| 8 | {"name": "filename", "type": "text", "label": "File"}, |
| 9 | {"name": "content_type", "type": "text", "label": "Type"}, |
| 10 | {"name": "size", "type": "integer", "default": "0", "label": "Bytes"}, |
| 11 | {"name": "content_sha256", "type": "text", "label": "Content Hash", |
| 12 | "help": "Of the stored bytes. The dedup key: photographing the same receipt twice, or emailing what somebody already snapped, must not become two expenses."}, |
| 13 | {"name": "file_id", "type": "text", "label": "Stored File", |
| 14 | "help": "Where the bytes live under the owner's files. The image IS the audit evidence; everything else here is derived from it."}, |
| 15 | {"name": "source", "type": "enum", "default": "web", "label": "Source", |
| 16 | "enum": ["web", "phone", "email", "mcp", "api"], |
| 17 | "help": "Where it came in from. Capture has to happen at the moment the paper exists -- a receipt that needs a laptop later is a receipt that is already lost."}, |
| 18 | {"name": "category_hint", "type": "enum", "default": "receipt", "label": "Kind", |
| 19 | "enum": ["receipt", "bill", "invoice", "statement", "document", "photo"], |
| 20 | "help": "What the person capturing it says it is. Cheap at capture time, expensive to infer afterwards."}, |
| 21 | {"name": "status", "type": "enum", "default": "pending", "label": "Status", |
| 22 | "enum": ["pending", "processing", "extracted", "confirmed", "ignored", "error"], |
| 23 | "transitions": { |
| 24 | "pending": ["processing", "extracted", "ignored", "error"], |
| 25 | "processing": ["extracted", "error", "ignored"], |
| 26 | "extracted": ["confirmed", "ignored", "error"], |
| 27 | "error": ["pending", "ignored"], |
| 28 | "ignored": ["pending"], |
| 29 | "confirmed": [] |
| 30 | }, |
| 31 | "help": "pending goes straight to extracted when a pass reads it in one go; processing exists for an engine slow enough that somebody would otherwise wonder whether it hung. error returns to pending so a fixed engine can retry the backlog. ignored is for the blurry duplicate nobody needs -- reversible, because 'ignore' should never mean 'lose'. confirmed is terminal: the record it produced is the thing to correct now, not this evidence of what arrived."}, |
| 32 | {"name": "ocr_text", "type": "textarea", "label": "Text Read", |
| 33 | "help": "What the engine read. Kept even when extraction found nothing, because it is searchable and because it is how somebody works out why a reading was wrong."}, |
| 34 | {"name": "ocr_engine", "type": "text", "label": "Engine", |
| 35 | "help": "Which engine produced the text and the guess. Recorded so a bad batch can be found and re-run when the engine changes."}, |
| 36 | {"name": "extracted", "type": "textarea", "label": "Extraction (JSON)", |
| 37 | "help": "The structured GUESS: vendor, date, total, tax, line items. Stored, shown beside the image, never posted."}, |
| 38 | {"name": "confidence", "type": "text", "default": "0", "label": "Confidence", |
| 39 | "help": "How much was actually found, not how sure a model claims to be. For sorting 'probably fine' above 'look at this one' -- never a threshold for acting unattended."}, |
| 40 | {"name": "confirmed_record", "type": "text", "label": "Became", |
| 41 | "help": "collection/id of what a human created from this. Present means confirmed: one scan, one record, however many times anyone clicks."}, |
| 42 | {"name": "error", "type": "text", "label": "Error"}, |
| 43 | {"name": "attempts", "type": "integer", "default": "0", "label": "Attempts", |
| 44 | "help": "Processing tries. A document that fails repeatedly stops being retried and starts being visible, rather than churning a paid extractor forever."}, |
| 45 | {"name": "notes", "type": "textarea", "label": "Notes"}, |
| 46 | {"name": "owner_id", "type": "text", "label": "Owner"}, |
| 47 | {"name": "entity_id", "type": "text", "label": "Entity", |
| 48 | "relation": {"collection": "entities", "display_field": "name"}}, |
| 49 | {"name": "created_at", "type": "datetime", "read_only": true, "label": "Received"} |
| 50 | ], |
| 51 | "hooks": {"before_write": "hook_scans"}, |
| 52 | "search": {"fields": ["filename", "ocr_text", "notes"]}, |
| 53 | "forms": {"default": {"fields": ["filename", "source", "category_hint", "notes"]}}, |
| 54 | "views": {"list_mode": "table", |
| 55 | "list_fields": ["created_at", "filename", "category_hint", "status", |
| 56 | "confidence"], |
| 57 | "filter_fields": ["status", "category_hint", "source"]} |
| 58 | } |