@talonic/docs 0.6.1 → 0.7.1

package/dist/seo.d.ts

@@ -549,7 +549,7 @@ declare const PLATFORM_FAQ: {
     question: string;
     answer: string;
 }[];
-declare const LLMS_TXT = "# Talonic\n\n> Extract any document into schema-validated data with a single API call.\n\nTalonic is an AI-powered document structuring platform. It ingests unstructured documents (PDFs, scans, images, spreadsheets), discovers fields through semantic clustering, and produces structured datasets with per-cell provenance, confidence scores, and reasoning traces.\n\n## Documentation\n\n- [API Documentation](https://talonic.com/docs): Complete REST API reference\n- [Platform Guide](https://talonic.com/docs/platform): Product documentation and feature guide\n- [OpenAPI Spec](https://talonic.com/docs/openapi.json): Machine-readable API specification\n\n## API Quick Start\n\nBase URL: https://api.talonic.com\nAuth: Bearer token (tlnc_ prefix)\n\nExtract a document:\n```\ncurl -X POST https://api.talonic.com/v1/extract \\\n  -H \"Authorization: Bearer tlnc_sk_live_...\" \\\n  -F \"file=@document.pdf\" \\\n  -F 'schema={\"vendor_name\":\"string\",\"total\":\"number\"}'\n```\n\n## Key Endpoints\n\n- POST /v1/extract \u2014 Extract structured data from a document\n- GET /v1/documents \u2014 List and manage documents\n- GET /v1/extractions \u2014 Query extraction results\n- POST /v1/schemas \u2014 Create reusable extraction schemas\n- GET /v1/jobs \u2014 Track async extraction jobs\n- POST /v1/sources \u2014 Manage document sources with API keys\n\n## Platform Features\n\n- 25+ file format support (PDF, DOCX, images, spreadsheets, archives)\n- Field Registry: unified knowledge graph of canonical fields with tier system\n- 4-phase extraction pipeline: Resolve \u2192 Agent \u2192 Validate \u2192 Re-read\n- Document linking and case discovery through shared entities\n- Validation checks, golden samples, and approval gates\n- Export to webhooks, REST APIs, SFTP, email, cloud storage\n- HMAC-SHA256 signed webhooks with retry policy\n\n## Optional\n\n- [API Documentation (Markdown)](https://talonic.com/docs/api.md)\n- [Platform Guide (Markdown)](https://talonic.com/docs/platform.md)\n";
+declare const LLMS_TXT = "# Talonic\n\n> AI-powered document structuring platform that turns unstructured files into schema-validated, provenance-tracked structured data.\n\nTalonic ingests documents in 25+ formats (PDFs, scans, images, spreadsheets, emails, archives), discovers every data field through AI extraction and semantic clustering, and produces structured datasets with per-cell confidence scores, reasoning traces, and source provenance. It runs a single deployable stack with Postgres + pgvector, Anthropic Claude for extraction, and Mistral Document AI for OCR.\n\n## How It Works\n\n1. **Upload** \u2014 Drag files/folders into Inputs or ingest via API. ZIP archives unpack automatically. Files are deduplicated via SHA-256 hashing.\n2. **Extract** \u2014 Each document goes through Document AI OCR (converts to Markdown), classification against a 529-type ontology, and AI field extraction (discovers every data point with confidence and source text).\n3. **Build Schema** \u2014 Extracted fields resolve into the Field Registry (canonical names, semantic clusters, master instructions). Define a user template selecting the fields you need.\n4. **Run Job** \u2014 A 4-phase pipeline fills every cell in a documents \u00D7 fields grid. ~30% filled instantly from graph matches, ~70% from AI agents.\n5. **Deliver** \u2014 Push approved data to webhooks, REST APIs, SFTP, email, or S3/R2 cloud storage.\n\n## Sources & Documents\n\nSupported formats across three processing paths:\n- **Text fast-path** (direct read): TXT, MD, HTML, XML, JSON, EML, CSV\n- **AI Vision** (multimodal): PNG, JPG, JPEG, GIF, WEBP\n- **OCR** (Mistral Document AI \u2192 Markdown): PDF, DOCX, DOC, PPTX, PPT, XLSX, XLS, XLSM, MSG, BMP\n- **Archives**: ZIP (recursive unpack)\n\nUpload methods: drag-and-drop UI, folder upload (preserves file paths), API upload (single, batch up to 200 files, or archive up to 500 MB). Batch mode available at 50% cost with 48-hour delivery window.\n\nEvery document is classified into a canonical type from the 529-type ontology (e.g., \"Employment Contract\", \"Invoice\", \"Bill of Lading\"). Classification is language-agnostic \u2014 a German Arbeitsvertrag maps to the same type as an English Employment Contract. Unresolvable documents get \"Unclassified Document\".\n\nDocument detail tabs: Raw Extraction (every field with confidence), Resolved Data (mapped to registry), Processing Log (per-stage timing), Original File.\n\n## Field Registry\n\nThe unified knowledge graph of all canonical fields, growing smarter with every document processed.\n\n- **Tier 1 (Core)**: Universal fields across many document types. Most reliable.\n- **Tier 2 (Established)**: Promoted from Tier 3 after frequency thresholds. Production-ready.\n- **Tier 3 (Emerging)**: Newly discovered from a few documents. May promote as more data arrives.\n\nFields with similar meanings cluster automatically via AI embeddings (e.g., \"Vendor Name\", \"Supplier Name\", \"Company Name\" \u2192 same cluster). Master instructions are AI-synthesized extraction directives that improve accuracy over time.\n\n## Schemas\n\nTwo types: **Generated schemas** (auto-created per document type from Tier 1+2 fields) and **User templates** (user-defined output structures).\n\nTemplate workflow: name it \u2192 add fields (display name, data type, extraction instructions) \u2192 map to registry (exact/semantic/composite matching) \u2192 add reference tables \u2192 publish an immutable version.\n\nField features: format constraints (regex validation with empty/flag/constant fallback), modifiers (date/number format, alias mapping, max_length), bypass strategies (constant, generator, reference lookup \u2014 skip LLM), capture submoves (match \u2192 compute \u2192 reason), output name remapping.\n\nVersioning: Live (published, read-only), Workshop (mutable draft), Version History (timeline with diff). Test extraction compares draft vs. live results before publishing.\n\n## Extraction Jobs (Runs)\n\nA job applies a schema to documents, producing a grid (rows = documents, columns = fields). Navigate to Structuring \u2192 Runs \u2192 New.\n\n**4-phase pipeline:**\n1. **Resolve** \u2014 ~30% of cells in seconds. Graph matches, fuzzy name matching, concept-synonym expansion, 3-tier reference lookup (normalize \u2192 fuzzy \u2192 AI), description scan. No AI calls (except rare Haiku fallback). Values normalized: dates \u2192 YYYY/MM/DD, numbers \u2192 2 decimal places.\n2. **Agent** \u2014 AI reviews gap patterns and produces typed strategy per field: compute (formula from grid values), transfer (copy from equivalent field), extract (re-read document with instructions, 5 concurrent), skip (with reasoning). Fields with manual instructions are always extracted, never skipped.\n3. **Validation** \u2014 Cross-field sanity checks: date_sanity, amount_mismatch, lookup_failed, low_confidence_outlier, unexpected_empty. Flags are informational only \u2014 never block output.\n4. **Re-read** \u2014 Context-aware gap filling. For each empty/low-confidence cell, AI re-reads the original document with field instruction + full grid context. Respects the confidence gate: cells \u2265 0.7 confidence are permanently protected.\n\nPer-cell provenance: confidence (0.0\u20131.0), resolution_type (graph_match | agent_derived | source_reread | unresolved), phase (1\u20134), reasoning trace, source reference (document, page, field).\n\n## Cases & Document Linking\n\nRegistry fields can be link keys: Identity (company/person names), Transaction (contract/PO/invoice numbers), Reference (project codes, cost centers). The linking pipeline normalizes values and builds a bipartite graph of documents \u2194 entities.\n\nA **case** = 2+ documents connected through transaction/reference entities. An **entity group** = 2+ documents connected through identity-only entities. High-frequency entities (>30% of documents) are auto-excluded from case formation.\n\nCase detail: documents, shared entities, evidence chain, timeline, AI-generated narration. Document Graph provides a D3-force visual layout. Case templates auto-discovered after 3+ cases form.\n\n## Smart Matching\n\nUpload CSV/Excel as reference datasets. Define field-to-field comparisons with weighted strategies: exact (case-insensitive), fuzzy (token-based with similarity threshold), date_range (configurable tolerance), numeric_range (percentage or absolute tolerance). AI can auto-suggest field mappings.\n\nResults: top 5 candidates per document with confidence scores and per-field evidence breakdown.\n\n## Validation & Quality\n\n- **Validation checks**: Schema-level rules (field format, value range, cross-field consistency, AI-proposed coherence). Run during Phase 3.\n- **Golden samples**: Manually-created reference datasets. Benchmark runs compare extraction vs. golden for per-field accuracy with AI judge verdicts.\n- **Approval gates**: Threshold-based auto-approve/flag (minimum confidence, validation pass rate, field coverage). Failed rows go to manual review queue.\n\n## Delivery\n\nDestinations: Webhook (HMAC-SHA256 signed), REST API (configurable headers), SFTP, Email (attachments), S3/R2 (cloud storage).\n\nField mappings transform output fields to match destination format. Triggers: auto on approval (stage/push), scheduled (cron), or manual. Failed exports retry with exponential backoff.\n\nDialects control serialization: date_format, number_locale, CSV delimiter, null representation, boolean format, encoding (UTF-8, UTF-8-BOM, ISO-8859-1).\n\n## Search & Navigation\n\n**Omnisearch** (Cmd+K / Ctrl+K): searches across documents, extracted values, field names, schema names, and sources simultaneously.\n\nDocument filters: field-value conditions with autocomplete, comparison operators (eq, contains, gt, between, is_empty), combinable. URL-serializable and saveable as presets.\n\nKeyboard shortcuts: Cmd+K (search), Cmd+J (quick extract), Escape (close overlays).\n\n## Team & Settings\n\n4 roles: Viewer (read-only), Member (full CRUD), Admin (+ team management), Owner (+ billing, API keys, org settings). New members auto-match by email domain with pending approval.\n\nUsage & Registry: per-feature cost breakdown, daily cost chart, call log with model/tokens/cost. Admin master view for cross-tenant stats.\n\n## API\n\nBase URL: `https://api.talonic.com`. Auth: `Authorization: Bearer tlnc_...` (SHA-256 hashed, shown once at creation). Scopes: extract, read, write.\n\nKey endpoints:\n- POST /v1/extract \u2014 Synchronous/async document extraction (`include_markdown=true` returns OCR text, `processing_mode=batch` for 50% cost)\n- GET /v1/documents \u2014 List with cursor pagination; GET /v1/documents/:id/markdown for OCR text\n- GET /v1/extractions \u2014 Query results and field corrections\n- POST /v1/schemas \u2014 Create/manage extraction schemas\n- GET /v1/jobs \u2014 Track async jobs and results\n- POST /v1/sources \u2014 Manage API sources and document ingest\n- POST /v1/webhooks \u2014 Configure webhook endpoints\n\nWebhook events: extraction.completed, job.completed, export.completed, validation.completed. All HMAC-SHA256 signed with retry on failure.\n\n## Agent\n\nThe embedded AI assistant accessible from any page. Two modes:\n- **Chat mode** \u2014 Ask questions about the platform, your documents, extraction results, schemas, or workflows. Grounded in platform documentation.\n- **Planning mode** \u2014 Request actions (create schemas, run jobs, configure exports). The agent builds a plan, confirms with you, then executes.\n\nDocument upload flow: Cmd+J or the upload button opens quick extract. Drop a file, select a schema (or let AI discover fields), and get structured results.\n\n## Documentation\n\n- [API Documentation](https://talonic.com/docs): Complete REST API reference\n- [Platform Guide](https://talonic.com/docs/platform): Product documentation and feature guide\n- [OpenAPI Spec](https://talonic.com/docs/openapi.json): Machine-readable API specification\n";
 declare const LLMS_FULL_TXT_HEADER = "# Talonic \u2014 Full Documentation\n\n> This file contains the complete Talonic documentation for LLM consumption.\n> For a summary, see llms.txt.\n\n";
 
 export { API_FAQ, API_NAV_SECTIONS, API_SECTION_META, LLMS_FULL_TXT_HEADER, LLMS_TXT, OPENAPI_SPEC, PLATFORM_FAQ, PLATFORM_NAV_SECTIONS, PLATFORM_SECTION_META, type SectionMeta };

package/dist/seo.js

@@ -292,52 +292,134 @@ var PLATFORM_FAQ = [
 ];
 var LLMS_TXT = `# Talonic
 
-> Extract any document into schema-validated data with a single API call.
+> AI-powered document structuring platform that turns unstructured files into schema-validated, provenance-tracked structured data.
 
-Talonic is an AI-powered document structuring platform. It ingests unstructured documents (PDFs, scans, images, spreadsheets), discovers fields through semantic clustering, and produces structured datasets with per-cell provenance, confidence scores, and reasoning traces.
+Talonic ingests documents in 25+ formats (PDFs, scans, images, spreadsheets, emails, archives), discovers every data field through AI extraction and semantic clustering, and produces structured datasets with per-cell confidence scores, reasoning traces, and source provenance. It runs a single deployable stack with Postgres + pgvector, Anthropic Claude for extraction, and Mistral Document AI for OCR.
 
-## Documentation
+## How It Works
 
-- [API Documentation](https://talonic.com/docs): Complete REST API reference
-- [Platform Guide](https://talonic.com/docs/platform): Product documentation and feature guide
-- [OpenAPI Spec](https://talonic.com/docs/openapi.json): Machine-readable API specification
+1. **Upload** \u2014 Drag files/folders into Inputs or ingest via API. ZIP archives unpack automatically. Files are deduplicated via SHA-256 hashing.
+2. **Extract** \u2014 Each document goes through Document AI OCR (converts to Markdown), classification against a 529-type ontology, and AI field extraction (discovers every data point with confidence and source text).
+3. **Build Schema** \u2014 Extracted fields resolve into the Field Registry (canonical names, semantic clusters, master instructions). Define a user template selecting the fields you need.
+4. **Run Job** \u2014 A 4-phase pipeline fills every cell in a documents \xD7 fields grid. ~30% filled instantly from graph matches, ~70% from AI agents.
+5. **Deliver** \u2014 Push approved data to webhooks, REST APIs, SFTP, email, or S3/R2 cloud storage.
+
+## Sources & Documents
+
+Supported formats across three processing paths:
+- **Text fast-path** (direct read): TXT, MD, HTML, XML, JSON, EML, CSV
+- **AI Vision** (multimodal): PNG, JPG, JPEG, GIF, WEBP
+- **OCR** (Mistral Document AI \u2192 Markdown): PDF, DOCX, DOC, PPTX, PPT, XLSX, XLS, XLSM, MSG, BMP
+- **Archives**: ZIP (recursive unpack)
+
+Upload methods: drag-and-drop UI, folder upload (preserves file paths), API upload (single, batch up to 200 files, or archive up to 500 MB). Batch mode available at 50% cost with 48-hour delivery window.
+
+Every document is classified into a canonical type from the 529-type ontology (e.g., "Employment Contract", "Invoice", "Bill of Lading"). Classification is language-agnostic \u2014 a German Arbeitsvertrag maps to the same type as an English Employment Contract. Unresolvable documents get "Unclassified Document".
+
+Document detail tabs: Raw Extraction (every field with confidence), Resolved Data (mapped to registry), Processing Log (per-stage timing), Original File.
+
+## Field Registry
+
+The unified knowledge graph of all canonical fields, growing smarter with every document processed.
+
+- **Tier 1 (Core)**: Universal fields across many document types. Most reliable.
+- **Tier 2 (Established)**: Promoted from Tier 3 after frequency thresholds. Production-ready.
+- **Tier 3 (Emerging)**: Newly discovered from a few documents. May promote as more data arrives.
+
+Fields with similar meanings cluster automatically via AI embeddings (e.g., "Vendor Name", "Supplier Name", "Company Name" \u2192 same cluster). Master instructions are AI-synthesized extraction directives that improve accuracy over time.
+
+## Schemas
+
+Two types: **Generated schemas** (auto-created per document type from Tier 1+2 fields) and **User templates** (user-defined output structures).
+
+Template workflow: name it \u2192 add fields (display name, data type, extraction instructions) \u2192 map to registry (exact/semantic/composite matching) \u2192 add reference tables \u2192 publish an immutable version.
+
+Field features: format constraints (regex validation with empty/flag/constant fallback), modifiers (date/number format, alias mapping, max_length), bypass strategies (constant, generator, reference lookup \u2014 skip LLM), capture submoves (match \u2192 compute \u2192 reason), output name remapping.
+
+Versioning: Live (published, read-only), Workshop (mutable draft), Version History (timeline with diff). Test extraction compares draft vs. live results before publishing.
+
+## Extraction Jobs (Runs)
+
+A job applies a schema to documents, producing a grid (rows = documents, columns = fields). Navigate to Structuring \u2192 Runs \u2192 New.
+
+**4-phase pipeline:**
+1. **Resolve** \u2014 ~30% of cells in seconds. Graph matches, fuzzy name matching, concept-synonym expansion, 3-tier reference lookup (normalize \u2192 fuzzy \u2192 AI), description scan. No AI calls (except rare Haiku fallback). Values normalized: dates \u2192 YYYY/MM/DD, numbers \u2192 2 decimal places.
+2. **Agent** \u2014 AI reviews gap patterns and produces typed strategy per field: compute (formula from grid values), transfer (copy from equivalent field), extract (re-read document with instructions, 5 concurrent), skip (with reasoning). Fields with manual instructions are always extracted, never skipped.
+3. **Validation** \u2014 Cross-field sanity checks: date_sanity, amount_mismatch, lookup_failed, low_confidence_outlier, unexpected_empty. Flags are informational only \u2014 never block output.
+4. **Re-read** \u2014 Context-aware gap filling. For each empty/low-confidence cell, AI re-reads the original document with field instruction + full grid context. Respects the confidence gate: cells \u2265 0.7 confidence are permanently protected.
+
+Per-cell provenance: confidence (0.0\u20131.0), resolution_type (graph_match | agent_derived | source_reread | unresolved), phase (1\u20134), reasoning trace, source reference (document, page, field).
+
+## Cases & Document Linking
 
-## API Quick Start
+Registry fields can be link keys: Identity (company/person names), Transaction (contract/PO/invoice numbers), Reference (project codes, cost centers). The linking pipeline normalizes values and builds a bipartite graph of documents \u2194 entities.
 
-Base URL: https://api.talonic.com
-Auth: Bearer token (tlnc_ prefix)
+A **case** = 2+ documents connected through transaction/reference entities. An **entity group** = 2+ documents connected through identity-only entities. High-frequency entities (>30% of documents) are auto-excluded from case formation.
 
-Extract a document:
-\`\`\`
-curl -X POST https://api.talonic.com/v1/extract \\
-  -H "Authorization: Bearer tlnc_sk_live_..." \\
-  -F "file=@document.pdf" \\
-  -F 'schema={"vendor_name":"string","total":"number"}'
-\`\`\`
+Case detail: documents, shared entities, evidence chain, timeline, AI-generated narration. Document Graph provides a D3-force visual layout. Case templates auto-discovered after 3+ cases form.
 
-## Key Endpoints
+## Smart Matching
 
-- POST /v1/extract \u2014 Extract structured data from a document
-- GET /v1/documents \u2014 List and manage documents
-- GET /v1/extractions \u2014 Query extraction results
-- POST /v1/schemas \u2014 Create reusable extraction schemas
-- GET /v1/jobs \u2014 Track async extraction jobs
-- POST /v1/sources \u2014 Manage document sources with API keys
+Upload CSV/Excel as reference datasets. Define field-to-field comparisons with weighted strategies: exact (case-insensitive), fuzzy (token-based with similarity threshold), date_range (configurable tolerance), numeric_range (percentage or absolute tolerance). AI can auto-suggest field mappings.
 
-## Platform Features
+Results: top 5 candidates per document with confidence scores and per-field evidence breakdown.
 
-- 25+ file format support (PDF, DOCX, images, spreadsheets, archives)
-- Field Registry: unified knowledge graph of canonical fields with tier system
-- 4-phase extraction pipeline: Resolve \u2192 Agent \u2192 Validate \u2192 Re-read
-- Document linking and case discovery through shared entities
-- Validation checks, golden samples, and approval gates
-- Export to webhooks, REST APIs, SFTP, email, cloud storage
-- HMAC-SHA256 signed webhooks with retry policy
+## Validation & Quality
 
-## Optional
+- **Validation checks**: Schema-level rules (field format, value range, cross-field consistency, AI-proposed coherence). Run during Phase 3.
+- **Golden samples**: Manually-created reference datasets. Benchmark runs compare extraction vs. golden for per-field accuracy with AI judge verdicts.
+- **Approval gates**: Threshold-based auto-approve/flag (minimum confidence, validation pass rate, field coverage). Failed rows go to manual review queue.
 
-- [API Documentation (Markdown)](https://talonic.com/docs/api.md)
-- [Platform Guide (Markdown)](https://talonic.com/docs/platform.md)
+## Delivery
+
+Destinations: Webhook (HMAC-SHA256 signed), REST API (configurable headers), SFTP, Email (attachments), S3/R2 (cloud storage).
+
+Field mappings transform output fields to match destination format. Triggers: auto on approval (stage/push), scheduled (cron), or manual. Failed exports retry with exponential backoff.
+
+Dialects control serialization: date_format, number_locale, CSV delimiter, null representation, boolean format, encoding (UTF-8, UTF-8-BOM, ISO-8859-1).
+
+## Search & Navigation
+
+**Omnisearch** (Cmd+K / Ctrl+K): searches across documents, extracted values, field names, schema names, and sources simultaneously.
+
+Document filters: field-value conditions with autocomplete, comparison operators (eq, contains, gt, between, is_empty), combinable. URL-serializable and saveable as presets.
+
+Keyboard shortcuts: Cmd+K (search), Cmd+J (quick extract), Escape (close overlays).
+
+## Team & Settings
+
+4 roles: Viewer (read-only), Member (full CRUD), Admin (+ team management), Owner (+ billing, API keys, org settings). New members auto-match by email domain with pending approval.
+
+Usage & Registry: per-feature cost breakdown, daily cost chart, call log with model/tokens/cost. Admin master view for cross-tenant stats.
+
+## API
+
+Base URL: \`https://api.talonic.com\`. Auth: \`Authorization: Bearer tlnc_...\` (SHA-256 hashed, shown once at creation). Scopes: extract, read, write.
+
+Key endpoints:
+- POST /v1/extract \u2014 Synchronous/async document extraction (\`include_markdown=true\` returns OCR text, \`processing_mode=batch\` for 50% cost)
+- GET /v1/documents \u2014 List with cursor pagination; GET /v1/documents/:id/markdown for OCR text
+- GET /v1/extractions \u2014 Query results and field corrections
+- POST /v1/schemas \u2014 Create/manage extraction schemas
+- GET /v1/jobs \u2014 Track async jobs and results
+- POST /v1/sources \u2014 Manage API sources and document ingest
+- POST /v1/webhooks \u2014 Configure webhook endpoints
+
+Webhook events: extraction.completed, job.completed, export.completed, validation.completed. All HMAC-SHA256 signed with retry on failure.
+
+## Agent
+
+The embedded AI assistant accessible from any page. Two modes:
+- **Chat mode** \u2014 Ask questions about the platform, your documents, extraction results, schemas, or workflows. Grounded in platform documentation.
+- **Planning mode** \u2014 Request actions (create schemas, run jobs, configure exports). The agent builds a plan, confirms with you, then executes.
+
+Document upload flow: Cmd+J or the upload button opens quick extract. Drop a file, select a schema (or let AI discover fields), and get structured results.
+
+## Documentation
+
+- [API Documentation](https://talonic.com/docs): Complete REST API reference
+- [Platform Guide](https://talonic.com/docs/platform): Product documentation and feature guide
+- [OpenAPI Spec](https://talonic.com/docs/openapi.json): Machine-readable API specification
 `;
 var LLMS_FULL_TXT_HEADER = `# Talonic \u2014 Full Documentation
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@talonic/docs",
-  "version": "0.6.1",
+  "version": "0.7.1",
   "description": "Talonic documentation components — API Reference & Platform Guide",
   "license": "UNLICENSED",
   "private": false,