DocVision AI

Displays the DocCanvas widget: Document panel (fields + color-coded confidence + provenance), Quality panel (heatmap + review queue), Versions, and Accept/Reject/Retry actions. Pass `document_id` (loads the persistent document) and/or `fields`.

docvision_doccanvas

Extracts pages (spec '1-3,7') into a NEW PDF (base64).

docvision_pages_extract

Extracts the real text of a PDF page by page WITH normalized positions (provenance). Flags pages with no text layer (scanned → raster OCR = worker layer). Fabricates no text.

docvision_extract_text

Extracts candidate fields (invoice/contract/form/statement/admin) from the provided pages, with PROVENANCE (page/area) and SUPPORTED_BY_SOURCE status. Absent values stay UNKNOWN.

docvision_extract_fields

Analyzes the provided integrity signals (metadata, per-page fonts, images, geometry, incremental updates) and flags ANOMALIES. NEVER concludes forgery (forgery verdict stays NOT_ESTABLISHED). Signature detected is not authenticated.

docvision_forensics_inspect

Opens a PDF (base64 or url) and returns pages, geometry, metadata and SHA-256 fingerprint. Real processing (pdf-lib).

docvision_inspect

Creates a relation between two documents (e.g. version/translation/source).

docvision_relation_add

Persisted annotations of a document.

docvision_annotations_list

Lists persisted batches.

docvision_batch_list

Lists persisted documents.

docvision_docs_list

Persisted extracted fields of a document.

docvision_fields_list

Patches of a document and their status.

docvision_patches_list

Merges several PDFs (list of base64/url sources) into one (base64), in the order provided.

docvision_merge

Resolves the text of ONE region from the outputs of several OCR engines (agreement/majority/divergence), with confidence and a review flag. Illegible → UNKNOWN (never invented). `critical:true` tightens the threshold (amounts/dates).

docvision_ocr_consensus

Compares two lists of page hashes/summaries (added/removed/unchanged).

docvision_page_diff

Applies the quality gate to each item of a batch and routes it → AUTO_APPROVED / REVIEW / FAILED, with metrics (including average confidence). FAILED/REVIEW go to the queue, not to export.

docvision_batch_process

Proposes a traceable edit (correct_ocr/rewrite/redact/delete…) WITHOUT touching the original. To be confirmed via apply.

docvision_patch_propose

Applies configurable quality thresholds to fields {name,type,value,confidence,disagreement?}. Blocks export if a critical field (amount/date/id) is below threshold. Default thresholds (amounts≥99.8, dates≥99…).

docvision_quality_gates

Consolidates gates + round-trip + review queue + forensics into one report, and highlights HIGH-CONFIDENCE errors (the most dangerous).

docvision_quality_report

Renders a PDF page as a PNG image (base64) — useful for preview, the DocCanvas, or before OCR. `scale` (default 2), `grayscale` for light restoration.

docvision_render_page

Metadata of a persistent document.

docvision_doc_get

Content of a specific version (or the current one) of a document.

docvision_version_get

REAL OCR of a PDF (including scanned): rasterizes each page and recognizes text with Tesseract, keeping per-word CONFIDENCE (words < 85% are flagged for review, never silently corrected). Execution runs in a CPU Vercel Sandbox (no serverless limit), so the model loads only once — no timeout. langs e.g. 'eng', 'fra', 'fra+eng'. Default 5 pages/call (max 20). If the Sandbox infra is not wired up, returns an honest WORKER_NOT_CONFIGURED status (never a fake result).

docvision_ocr_run

Produces a REAL EDITABLE .docx binary (Word styles, lists, tables) from a structure (blocks, or `pages` from docvision_structure_analyze). Returns the .docx as base64. Truly editable, not an image + text.

docvision_reconstruct_docx

Persistently records a field (value + provenance + confidence + status). Absent value → UNKNOWN.

docvision_field_record

Detects sensitive candidates (PII, financial, identifiers) in the provided pages, with provenance. Custom rules possible.

docvision_redaction_detect

Produces a redaction PLAN from the candidates + policy. PLAN only: the real burn-in (removal of the underlying content) = worker layer.

docvision_redaction_plan

Rejects (or retries) a patch without modifying the original.

docvision_patch_reject

Reorders pages according to a complete permutation (e.g. [3,1,2]) and returns a NEW PDF (base64).

docvision_pages_reorder

Records a review decision (ACCEPT/REJECT/RETRY) on a target (patch, region, field).

docvision_review_decide

Builds the review queue from resolved regions (or {engines}): keeps only what needs review, with the heatmap (green/yellow/red/gray).

docvision_review_queue

Rotates pages of a PDF (spec '1,3-5' or empty=all) and returns a NEW PDF (base64). The original is never modified.

docvision_pages_rotate

Compares the ORIGINAL and the GENERATED rendering (text/tables/images/critical fields) and returns scores + review areas. DocVision checks its own work (the gap is not hidden).

docvision_roundtrip_validate

Persistently saves a quality report for a document.

docvision_quality_save

Full-text search within a PDF with location (page + snippet). Real.

docvision_search

Compares two texts/versions: separates CONTENT CHANGE from visual MOVE, and flags changed NUMBERS (amounts/dates/IDs).

docvision_diff

Returns the default RBAC schema (viewer/reviewer/editor/admin/auditor roles, rules, approvals).

docvision_policy_show

Splits a PDF into two parts (after page `at`, default middle). Returns two PDFs (base64).

docvision_split

Transforms extracted text (pages) into a document tree (headings/paragraphs/lists). Honest heuristic; fine tables/figures via provided regions or worker.

docvision_structure_analyze

Versions of a document (original/current/derivations).

docvision_doc_history