Capabilities

Purple8 DocIntel handles the full document intelligence lifecycle — from raw bytes to a structured knowledge graph — across every common enterprise format.

Supported formats

Purple8 DocIntel natively parses 70+ document formats across 6 categories — no pre-conversion required.

Office & productivity

Format	Extension(s)	Notes
PDF	.pdf	Text extraction + OCR fallback for scanned pages
Word	.docx, .doc, .odt	Paragraph, heading, table structure preserved
Excel	.xlsx, .xls, .ods	All sheets; cell values + types
PowerPoint	.pptx, .ppt, .odp	Slide text, notes, shape labels
RTF	.rtf	Rich text formatting stripped to plain
EPUB	.epub	Spine chapters extracted as ordered pages

Web & text

Format	Extension(s)	Notes
HTML	.html, .htm	Tags stripped; structure preserved via headings
Plain text	.txt, .md, .markdown, .rst	Decoded UTF-8, latin-1 fallback
Email	.eml, .msg	Headers + body; attachments enumerated
CSV / TSV	.csv, .tsv	Rendered as pipe-delimited markdown tables
JSON	.json	Pretty-printed as structured text
XML	.xml	Tags preserved as structured context

Images (OCR)

Format	Extension(s)	Engine
Raster images	.png, .jpg, .jpeg, .tiff, .bmp, .webp	Tesseract · Azure Document Intelligence · Google Cloud Vision · AWS Textract
Sketch / whiteboard	.sketch	GPT-5 vision · Azure AI Vision 4.0 · Google Cloud Vision

CAD & BIM

Format	Extension(s)	Extracted content
IFC	.ifc, .ifczip	Project, site, building, storeys, spaces, IfcElements, materials, property sets, quantities, type objects
DXF	.dxf	Layers, entities, blocks, dimensions, text
DWG	.dwg	Via LibreDWG conversion to DXF
STEP	.step, .stp	Assembly tree, part names, B-rep metadata
STL	.stl	Triangle count, bounding box, part name
OBJ	.obj	Mesh groups, material references
GLTF / GLB	.gltf, .glb	Scene hierarchy, mesh names, materials
G-code	.gcode	Tool paths, layer count, estimated print time

Code & config

.py, .js, .ts, .css, .sql, .sh, .env, .yaml, .yml, .toml, .ini, .cfg — extracted as structured text with language metadata.

SAP

Format	Extension(s)	Notes
IDoc	.idoc	Fixed-width 1000-byte record format; segment tree reconstructed
ABAP source	.abap	Extracted as plain text with structure markers

LLM providers

Configure via LLM__PROVIDER environment variable:

Provider	Value	Notes
OpenAI	openai	Default; uses ModelRegistry tiers
Azure OpenAI	azure	Set LLM__AZURE_ENDPOINT + LLM__AZURE_DEPLOYMENT
Anthropic	anthropic	Claude 3.x via tool-use mode
Google Gemini	gemini	Via Vertex AI or AI Studio
Ollama	ollama	Set LLM__BASE_URL; air-gapped deployments

Model tiers (ModelRegistry)

Tier	Current model	Used for
FLAGSHIP	gpt-5.6-sol	AEC, Legal/Contract
DEFAULT	gpt-5.6-terra	Financial, Scientific, General
VISION	gpt-5.6-sol	Sketch / image parsing
NANO	gpt-6-luna	High-volume, cost-sensitive extraction

All four constants live in config.ModelRegistry. Upgrading models = changing four lines.

OCR engines

Engine	OCR__ENGINE value	Notes
Tesseract	tesseract	Default; local, no API key needed
Azure Document Intelligence	azure_doc_intel	Best for structured forms + tables
Google Cloud Vision	google_vision	Strong multilingual support
AWS Textract	textract	Best for US healthcare + financial forms

Domain extraction profiles

See Extraction Profiles → for full ontology details.

Profile	profile= value	Model tier	Chunking	Confidence
AEC / BIM	aec	FLAGSHIP	table_aware	≥ 0.75
Legal / Contract	contract	FLAGSHIP	clause	≥ 0.75
Financial / ERP	financial	DEFAULT	table_aware	≥ 0.68
Scientific	scientific	DEFAULT	section	≥ 0.70
General	general	DEFAULT	paragraph	≥ 0.65

Source connectors

See Connectors → for configuration details.

Connector	Auth	Protocol
Microsoft SharePoint	Azure AD service principal	Batch pull + change webhooks
Atlassian Confluence	API token	Batch pull by space key
AWS S3	IAM credentials / instance profile	Batch pull with prefix + glob filter
Generic webhook	Bearer token (header)	POST file bytes or URL

Job persistence

Jobs and SHA-256 content hashes are persisted in an embedded durable p8l log (SessionContentStore) at /data/docintel_jobs.p8l by default — zero external Redis or SQLite dependencies required. Child CPU workers use lazy initialisation (_LazySessionContentStore) so multi-core worker pools spawn with zero startup replay overhead.

Confidence calibration v0.12.5

LLMs systematically produce overconfident scores (clustered 0.85–1.0). DocIntel applies temperature scaling (T = 2.0) to spread the distribution via a logit → scale → sigmoid transform:

Raw LLM score	Calibrated score
0.95	~0.82
0.90	~0.74
0.80	~0.62
0.70	~0.58
0.50	0.50 (fixed point)

Applied after fuzzy dedup, before the per-profile confidence threshold. Invoice ECE improved from 0.30 → 0.13.

Read more →

Table extraction v0.6.7

Structured tables from PDFs, DOCX, and XLSX are no longer flattened into running text:

1. _tables_to_markdown() converts ParsedDocument.raw_tables into markdown table strings
2. _inject_table_windows() creates dedicated LLM context windows from those tables
3. Large tables are auto-split by row with header preservation
4. Each window is tagged with is_table=True metadata

This ensures line items, amounts, dates, and columnar data in tabular form get full extraction attention.

Read more →

Incremental processing v0.6.7

Resubmitting the same document (via any connector or API) costs zero processing:

1. SHA-256 content_hash computed from raw document bytes
2. All three job stores (memory, redis, sqlite) support get_by_hash() lookup
3. If a completed job with the same hash exists, the cached ExtractionResult is returned immediately
4. The new job is marked completed with a pipeline.incremental_skip log entry

Read more →

Multi-language support v0.12.5

Automatic language detection across 75+ languages via Lingua, paired with rich per-language extraction conventions:

Language	Script	Key guidance
German	Latin	Compound nouns split; formal Sie/du handling
French	Latin	Accented characters preserved; date format DD/MM/YYYY
Chinese	CJK	Simplified & traditional; no spaces between characters
Japanese	CJK + Kana	Kanji names with furigana; honorifics preserved
Korean	Hangul	Hangul + Hanja; family-name-first convention
Arabic	Arabic	Right-to-left; Hijri date normalisation
Hindi	Devanagari	Devanagari + Latin transliteration; honorifics
+ 8 more	Various	Spanish, Portuguese, Italian, Dutch, Russian, Turkish, Polish, Swedish

All entity names are extracted in native script with optional romanisation. JSON keys always returned in English.

Read more →