129 lines
11 KiB
Markdown
129 lines
11 KiB
Markdown
# Changelog
|
|
|
|
## 2026-01-18 - 1.11.0 - feat(vision)
|
|
process pages separately and make Qwen3-VL vision extraction more robust; add per-page parsing, safer JSON handling, reduced token usage, and multi-query invoice extraction
|
|
|
|
- Bank statements: split extraction into extractTransactionsFromPage and sequentially process pages to avoid thinking-token exhaustion
|
|
- Bank statements: reduced num_predict from 8000 to 4000, send single image per request, added per-page logging and non-throwing handling for empty or non-JSON responses
|
|
- Bank statements: catch JSON.parse errors and return empty array instead of throwing
|
|
- Invoices: introduced queryField to request single values and perform multiple simple queries (reduces model thinking usage)
|
|
- Invoices: reduced num_predict for invoice queries from 4000 to 500 and parse amounts robustly (handles European formats like 1.234,56)
|
|
- Invoices: normalize currency to uppercase 3-letter code, return safe defaults (empty strings / 0) instead of nulls, and parse net/vat/total with fallbacks
|
|
- General: simplified Ollama API error messages to avoid including response body content in thrown errors
|
|
|
|
## 2026-01-18 - 1.10.1 - fix(tests)
|
|
improve Qwen3-VL invoice extraction test by switching to non-stream API, adding model availability/pull checks, simplifying response parsing, and tightening model options
|
|
|
|
- Replaced streaming reader logic with direct JSON parsing of the /api/chat response
|
|
- Added ensureQwen3Vl() to check and pull the Qwen3-VL:8b model from Ollama
|
|
- Switched to ensureMiniCpm() to verify Ollama service is running before model checks
|
|
- Use /no_think prompt for direct JSON output and set temperature to 0.0 and num_predict to 512
|
|
- Removed retry loop and streaming parsing; improved error messages to include response body
|
|
- Updated logging and test setup messages for clarity
|
|
|
|
## 2026-01-18 - 1.10.0 - feat(vision)
|
|
add Qwen3-VL vision model support with Dockerfile and tests; improve invoice OCR conversion and prompts; simplify extraction flow by removing consensus voting
|
|
|
|
- Add Dockerfile_qwen3vl to provide an Ollama-based image for Qwen3-VL and expose the Ollama API on port 11434
|
|
- Introduce test/test.invoices.qwen3vl.ts and ensureQwen3Vl() helper to pull and test qwen3-vl:8b
|
|
- Improve PDF->PNG conversion and prompt in ministral3 tests (higher DPI, max quality, sharpen) and increase num_predict from 512 to 1024
|
|
- Simplify extraction pipeline: remove consensus voting, log single-pass results, and simplify OCR HTML sanitization/truncation logic
|
|
|
|
## 2026-01-18 - 1.9.0 - feat(tests)
|
|
add Ministral 3 vision tests and improve invoice extraction pipeline to use Ollama chat schema, sanitization, and multi-page support
|
|
|
|
- Add new vision-based test suites for Ministral 3: test/test.invoices.ministral3.ts and test/test.bankstatements.ministral3.ts (model ministral-3:8b).
|
|
- Introduce ensureMinistral3() helper to start/check Ollama/MiniCPM model in test/helpers/docker.ts.
|
|
- Switch invoice extraction to use Ollama /api/chat with a JSON schema (format) and streaming support (reads message.content).
|
|
- Improve HTML handling: sanitizeHtml() to remove OCR artifacts, concatenate multi-page HTML with page markers, and increase truncation limits.
|
|
- Enhance response parsing: strip Markdown code fences, robustly locate JSON object boundaries, and provide clearer JSON parse errors.
|
|
- Add PDF->PNG conversion (ImageMagick) and direct image-based extraction flow for vision model tests.
|
|
|
|
## 2026-01-18 - 1.8.0 - feat(paddleocr-vl)
|
|
add structured HTML output and table parsing for PaddleOCR-VL, update API, tests, and README
|
|
|
|
- Add result_to_html(), parse_markdown_table(), and parse_paddleocr_table() to emit semantic HTML and convert OCR/markdown tables to proper <table> elements
|
|
- Enhance result_to_markdown() with positional/type hints (header/footer/title/table/figure) to improve downstream LLM processing
|
|
- Expose 'html' in supported formats and handle output_format='html' in parse endpoints and CLI flow
|
|
- Update tests to request HTML output and extract invoice fields from structured HTML (test/test.invoices.paddleocr-vl.ts)
|
|
- Refresh README with usage, new images/tags, architecture notes, and troubleshooting for the updated pipeline
|
|
|
|
## 2026-01-17 - 1.7.1 - fix(docker)
|
|
standardize Dockerfile and entrypoint filenames; add GPU-specific Dockerfiles and update build and test references
|
|
|
|
- Added Dockerfile_minicpm45v_gpu and image_support_files/minicpm45v_entrypoint.sh; removed the old Dockerfile_minicpm45v and docker-entrypoint.sh
|
|
- Renamed and simplified PaddleOCR entrypoint to image_support_files/paddleocr_vl_entrypoint.sh and updated CPU/GPU Dockerfile references
|
|
- Updated build-images.sh to use *_gpu Dockerfiles and clarified PaddleOCR GPU build log
|
|
- Updated test/helpers/docker.ts to point to Dockerfile_minicpm45v_gpu so tests build the GPU variant
|
|
|
|
## 2026-01-17 - 1.7.0 - feat(tests)
|
|
use Qwen2.5 (Ollama) for invoice extraction tests and add helpers for model management; normalize dates and coerce numeric fields
|
|
|
|
- Added ensureOllamaModel and ensureQwen25 test helpers to pull/check Ollama models via localhost:11434
|
|
- Updated invoices test to use qwen2.5:7b instead of MiniCPM and removed image payload from the text-only extraction step
|
|
- Increased Markdown truncate limit from 8000 to 12000 and reduced model num_predict from 2048 to 512
|
|
- Rewrote extraction prompt to require strict JSON output and added post-processing to parse/convert numeric fields
|
|
- Added normalizeDate and improved compareInvoice to normalize dates and handle numeric formatting/tolerance
|
|
- Updated test setup to ensure Qwen2.5 is available and adjusted logging/messages to reflect the Qwen2.5-based workflow
|
|
|
|
## 2026-01-17 - 1.6.0 - feat(paddleocr-vl)
|
|
add PaddleOCR-VL full pipeline Docker image and API server, plus integration tests and docker helpers
|
|
|
|
- Add Dockerfile_paddleocr_vl_full and entrypoint script to build a GPU-enabled image with PP-DocLayoutV2 + PaddleOCR-VL and a FastAPI server
|
|
- Introduce image_support_files/paddleocr_vl_full_server.py implementing the full pipeline API (/parse, OpenAI-compatible /v1/chat/completions) and a /formats endpoint
|
|
- Improve image handling: decode_image supports data URLs, HTTP(S), raw base64 and file paths; add optimize_image_resolution to auto-scale images into the recommended 1080-2048px range
|
|
- Add test helpers (test/helpers/docker.ts) to build/start/health-check Docker images and new ensurePaddleOcrVlFull workflow
|
|
- Add comprehensive integration tests for bank statements and invoices (MiniCPM and PaddleOCR-VL variants) and update tests to ensure required containers are running before tests
|
|
- Switch MiniCPM model references to 'minicpm-v:latest' and increase health/timeout expectations for the full pipeline
|
|
|
|
## 2026-01-17 - 1.5.0 - feat(paddleocr-vl)
|
|
add PaddleOCR-VL GPU Dockerfile, pin vllm, update CPU image deps, and improve entrypoint and tests
|
|
|
|
- Add a new GPU Dockerfile for PaddleOCR-VL (transformers-based) with CUDA support, healthcheck, and entrypoint.
|
|
- Pin vllm to 0.11.1 in Dockerfile_paddleocr_vl to use the first stable release with PaddleOCR-VL support.
|
|
- Update CPU image: add torchvision==0.20.1 and extra Python deps (protobuf, sentencepiece, einops) required by the transformers-based server.
|
|
- Rewrite paddleocr-vl-entrypoint.sh to build vllm args array, add MAX_MODEL_LEN and ENFORCE_EAGER env vars, include --limit-mm-per-prompt and optional --enforce-eager, and switch to exec vllm with constructed args.
|
|
- Update tests to use the OpenAI-compatible PaddleOCR-VL chat completions API (/v1/chat/completions) with image+text message payload and model 'paddleocr-vl'.
|
|
- Add @types/node to package.json dependencies and tidy devDependencies ordering.
|
|
|
|
## 2026-01-16 - 1.4.0 - feat(invoices)
|
|
add hybrid OCR + vision invoice/document parsing with PaddleOCR, consensus voting, and prompt/test refactors
|
|
|
|
- Add hybrid pipeline documentation and examples (PaddleOCR + MiniCPM-V) and architecture diagram in recipes/document.md
|
|
- Integrate PaddleOCR: new OCR extraction functions and OCR-only prompt flow in test/test.node.ts
|
|
- Add consensus voting and parallel-pass optimization to improve reliability (multiple passes, hashing, and majority voting)
|
|
- Refactor prompts and tests: introduce /nothink token, OCR truncation limits, separate visual and OCR-only prompts, and improved prompt building in test/test.invoices.ts
|
|
- Update image conversion defaults (200 DPI, filename change) and add TypeScript helper functions for extraction and consensus handling
|
|
|
|
## 2026-01-16 - 1.3.0 - feat(paddleocr)
|
|
add PaddleOCR OCR service (Docker images, server, tests, docs) and CI workflows
|
|
|
|
- Add GPU and CPU PaddleOCR Dockerfiles; pin paddlepaddle/paddle and paddleocr to stable 2.x and install libgomp1 for CPU builds
|
|
- Avoid pre-downloading OCR models at build-time to prevent build-time segfaults; models are downloaded on first run
|
|
- Refactor PaddleOCR FastAPI server: respect CUDA_VISIBLE_DEVICES, support per-request language, cache default language instance and create temporary instances for other languages
|
|
- Add comprehensive tests (test.paddleocr.ts) and improve invoice extraction tests (parallelize passes, JSON OCR API usage, prioritize certain test cases)
|
|
- Add Gitea CI workflows for tag and non-tag Docker runs and release pipeline (docker build/push, metadata trigger)
|
|
- Update documentation (readme.hints.md) with PaddleOCR usage and add docker registry entry to npmextra.json
|
|
|
|
## 2026-01-16 - 1.2.0 - feat(paddleocr)
|
|
add PaddleOCR support: Docker images, FastAPI server, entrypoint and tests
|
|
|
|
- Add PaddleOCR FastAPI server implementation at image_support_files/paddleocr_server.py
|
|
- Remove old image_support_files/paddleocr-server.py and update entrypoint to import paddleocr_server:app
|
|
- Extend build-images.sh to build paddleocr (GPU) and paddleocr-cpu images and list them
|
|
- Extend test-images.sh to add paddleocr health/OCR tests, new test_paddleocr_image function, port config, and cleanup; rename test_image -> test_minicpm_image
|
|
|
|
## 2026-01-16 - 1.1.0 - feat(ocr)
|
|
add PaddleOCR GPU Docker image and FastAPI OCR server with entrypoint; implement OCR endpoints and consensus extraction testing
|
|
|
|
- Add Dockerfile_paddleocr for GPU-accelerated PaddleOCR image (pre-downloads PP-OCRv4 models, exposes port 5000, healthcheck, entrypoint)
|
|
- Add image_support_files/paddleocr-server.py: FastAPI app providing /ocr (base64), /ocr/upload (file), and /health endpoints; model warm-up on startup; structured JSON responses and error handling
|
|
- Add image_support_files/paddleocr-entrypoint.sh to configure environment, detect GPU/CPU mode, and launch uvicorn
|
|
- Update test/test.node.ts to replace streaming extraction with a consensus-based extraction flow (multiple passes, hashing of results, majority voting) and improve logging/prompt text
|
|
- Add test/test.invoices.ts: integration tests for invoice extraction that call PaddleOCR, build prompts with optional OCR text, run consensus extraction, and produce a summary report
|
|
|
|
## 2026-01-16 - 1.0.0 - initial release
|
|
Initial project files added with two small follow-up updates.
|
|
|
|
- initial: base project commit.
|
|
- update: two minor follow-up updates refining the initial commit. |