ScriptAIX

ScriptAIX is a multi-tenant, AI-powered transcription and document analysis platform engineered specifically to process, transcribe, structure, and analyze handwritten historical manuscripts and archival records. Designed as an independent, comprehensive pipeline, it bridges raw digitized scans and modern archival data formats, providing an end-to-end workflow from image ingestion to standardized XML and knowledge graph generation.

ScriptAIX accommodates multiple ingestion workflows tailored for archival collections and everyday transcription needs:

• Local Upload & Batch Import: Supports raw images (JPG, JPEG, PNG, TIFF, WebP) and multi-page documents (PDF, multi-frame TIFF). Multi-page files are automatically segmented into individual pages for processing.
• IIIF Manifest Ingestion: Ingests digital items directly via IIIF Manifest URLs (v2.1 and v3.0 APIs), downloading canvases in full resolution along with original metadata and attribution.
• Automated Email Ingestion: Includes an IMAP/SMTP mail agent that monitors an inbox, processes image/PDF attachments from whitelisted senders, and returns transcriptions as email text, Markdown (.md), and PDF attachments.
• Spool Directory Processing: Implements a local spooler queue that handles background job queues, resubmissions, and offline job orchestration.

The transcription engine combines modern LLM APIs with paleographical prompting strategies:

• Specialized Paleography Presets: Offers prompt presets tailored to specific scripts, eras, and record classes (e.g., 19th-century Dutch records, Sütterlin, church registers, notarial records).
• Diplomatic vs. Normalized Transcriptions: Generates strict diplomatic transcriptions (preserving original spelling, line breaks, abbreviations, stamps, margins) alongside optional normalized reading versions.
• Cognitive Chain-of-Thought: Models are prompted to generate an explicit <analyse> block to mentally map out layout and paleographical challenges before beginning the transcription.
• Model Routing & Caching: Utilizes Claude models (Opus, Fable, Sonnet, Haiku) with system prompt ephemeral caching to reduce repeated token costs. Extensible to OpenAI-compatible APIs.
• Multi-Page Context Continuation: Supports chained page context where the transcription output of the previous page is passed forward to maintain sentence continuity across breaks and preserve consistent naming.
• Anthropic Message Batches API: Allows submission of bulk transcription jobs via Anthropic's Batch API (providing a 50% discount on token pricing), tracked asynchronously with automated retrieval.

Beyond text transcription, ScriptAIX performs physical layout segmentation and alignment:

• Baseline & Polygon Segmentation: Runs baseline layout segmentation to detect text regions, line boundaries, and coordinate polygons.
• Column Clustering: Analyzes horizontal span overlaps to identify multi-column layouts and register gutters without collapsing columns, sorting lines top-to-bottom per column.
• Interactive Overlays: Visualizes text line polygons as SVG overlays superimposed directly on the scan inside the viewer. Hovering over a polygon highlights the corresponding transcription text, and vice versa.
• ALTO & PageXML Export: Generates standardized ALTO v4 XML and PageXML files with word and line bounding boxes.
• METS Wrapper Generation: Produces METS XML files linking the physical scan, ALTO, PageXML, and Markdown representations.

ScriptAIX transforms unstructured transcripts into structured archival metadata and relational graphs:

• Automated Entity Extraction (NER): Extracts structured metadata directly from transcriptions into normalized JSON/YAML front-matter, identifying document dates, locations, persons (including maiden names and roles), relationships, and vital events.
• Interactive Relationship Graphs: Automatically renders extracted kinship and professional networks as graphs, featuring deep-zoom pan capabilities and high-resolution SVG/PNG downloads.
• Unified Knowledge Base: Maintains a centralized register of persons, places, events, and relationships synchronized across all tenant documents.
• Canonical Disambiguation: Allows researchers to merge duplicate person/place entries across documents while retaining all supporting source citations and variant spellings.

The browser-based workbench provides extensive inspection, correction, and curation tools:

• Dual-Pane Synchronized Workspace: Displays the original scan alongside the transcription with independent zooming, panning, 90° rotation, color inversion, and brightness/contrast adjustments.
• Magnifying Loupe: Floating loupe (up to 8× magnification) with rotation-aware coordinate translation for inspecting ink strokes and difficult hands.
• Revision History & Visual Diff: Compares the active text against .backup snapshots using a visual line-by-line and inline word-diff highlighting additions and deletions.
• Comprehensive Search Engine: Features Lexical search, Phonetic Soundex matching, Regex/Wildcard search, and Multilingual Hybrid Semantic Search (vector-based meaning search combined with FTS).
• Quality Workflow: Three-tier review pipeline (Draft → Reviewed → Released) with version locking, comment logging, and review permissions.
• Audit Trail: Cryptographically verifiable audit log with SHA-256 hash chaining to track all document edits, renames, status updates, and tenant transfers.

ScriptAIX exports data into an extensive set of scholarly, archival, genealogical, and office formats:

• Multi-Tenant Workspace Isolation: Segregates document repositories, search indexes, knowledge bases, and SQL records per tenant workspace.
• Role-Based Access Control (RBAC): Supports granular roles (Viewer, Editor, Admin, Superuser) controlling upload, edit, review, release, and export capabilities.
• Authentication & 2FA: JWT-based session tokens with TOTP Two-Factor Authentication and password recovery workflows.
• Token Accounting: Tracks API token consumption (Input, Output, Cache-Read, Cache-Write) with cost calculation and CSV reporting per user, tenant, and job.
• Tenant File Transfer: Admin tooling to safely move or copy complete transcription sets (Markdown, backups, overlays, and scans) between workspaces with conflict resolution policies.

Despite its capabilities, ScriptAIX should always be regarded as a research assistant rather than an authority. AI can misread names, dates, places, numbers and monetary amounts, particularly when the original manuscript is damaged or ambiguous. For that reason, ScriptAIX is designed to be transparent about uncertainty. Doubtful readings are explicitly marked, and illegible passages remain identified as such rather than silently "corrected".

Interested in testing? Email Erik for a link to the registration page.