HKUDS/RAG-Anything
62.5
Adequate · 4 November 2026
20.7k
lines of production code
Python
primary language
4
measurements over time
What this system is
Features
Added reproduce directory with RAGAnything integration examples and evaluation tools
The new \reproduce\ directory provides example scripts and configuration files to demonstrate integrating MinerU parsing with RAGAnything. It includes \index.py\ and \query.py\ for processing documents and performing text or multimodal queries, an \env.example\ file for configuring LLM, embedding, and reranking bindings, and \llm\_answer\_evaluator.py\ for assessing RAG output accuracy and quality.
reproduce · high confidence
Added script to cache tiktoken models for offline use
A new script, create\_tiktoken\_cache.py, has been added to pre-download and cache the o200k\_base and cl100k\_base tiktoken encodings. This ensures that the RAG-Anything application can function in offline environments by providing the necessary tokenizer data locally, rather than attempting to download it at runtime.
scripts · high confidence
New and updated example scripts for batch processing, multimodal insertion, and local LLM integrations
The examples directory now includes several new demonstration scripts: batch\_dry\_run\_example.py and batch\_processing\_example.py show how to use the BatchParser for scanning and processing multiple documents in parallel; insert\_content\_list\_example.py demonstrates direct insertion of structured content (text, images, tables, equations) into the knowledge base; modalprocessors\_example.py illustrates using image, table, and equation processors directly; image\_format\_test.py and office\_document\_test.py provide testing scripts for image and Office document parsing via MinerU (with LibreOffice checks); enhanced\_markdown\_example.py covers markdown-to-PDF conversion with WeasyPrint and Pandoc backends. Additionally, integration examples for LM Studio, MiniMax, and Ollama have been added or updated, showing how to configure local and cloud LLM/embedding endpoints, handle model discovery failures, and manage API keys via environment variables.
examples · high confidence
RAGAnything v1.4.1 introduces modular architecture with batch processing, callbacks, and multimodal support
RAGAnything has been restructured into a modular package (v1.4.1) that adds significant new capabilities while maintaining backward compatibility. Users can now process multiple documents in parallel using the new \BatchMixin\ and \BatchParser\ classes, which support progress tracking, timeouts, and dry-run modes. A new event-driven callback system (\ProcessingCallback\, \CallbackManager\) allows users to hook into parsing, insertion, and query stages for monitoring and custom logic. The library now supports audio and video modal processing via optional \AudioModalProcessor\ and \VideoModalProcessor\ classes (requiring \faster-whisper\ and \scenedetect\ respectively). Configuration is centralized in \RAGAnythingConfig\ with environment variable support, and the package exposes a public media URL mapping system (\asset\_urls.py\) for serving local assets via CDN. The core \Parser\ class and resilience utilities (retry, circuit breaker) are also exposed, with optional features like custom parser registration and multilingual prompt management available when installed.
raganything · high confidence
Test coverage
Added test coverage for public media URL mapping, audio processing, callbacks, and core modules
This change adds a suite of new tests to the \tests/\ directory. It introduces \test\_asset\_urls.py\ to verify the public URL mapping logic for local media paths, including environment variable handling and special character encoding. New tests for the audio modal processor (\test\_audio\_processor.py\) cover file detection, initialization, and transcription formatting. The callback system is tested (\test\_callbacks.py\) to ensure event dispatching, metrics collection, and error handling work correctly. Additionally, core module tests (\test\_core\_modules.py\) validate configuration defaults, content separation, and image encoding, while other new tests cover Chinese CID font registration, event loop closure, content list alias handling, custom parser plugins, document status creation, and force multimodal reprocessing. Also in this area: 1 commit adding/updating tests in tests; 20 commits adding/updating tests in tests; Added tests for batch parsing, document context isolation, and multimodal content handling; Added tests for image validation and PaddleOCR metadata exclusion; Added regression test suites for parser, processor, and concurrency fixes.
tests · high confidence · unverified
Dependencies
Introduce pyproject.toml and pin dependency versions
The project now uses pyproject.toml as the primary build configuration, defining the package name 'raganything' and its core dependencies: huggingface\_hub, lightrag-hku (pinned below 1.5), mineru\[core\] (pinned to \>=3.4.1), and tqdm. This change also introduces optional dependency groups for image, text, office, paddleocr, markdown, audio, video, and all features, alongside dev dependencies for testing and linting. The requirements.txt file is updated to reflect these core dependencies and includes comments explaining the mineru version constraint to avoid compatibility issues with pdftext.
(dependencies) · high confidence
Housekeeping
Initial project scaffolding and documentation
The repository has been initialized with the core project structure, including the Python package definition in setup.py, environment configuration in env.example, and comprehensive English and Chinese documentation in README.md and README\_zh.md. This initial commit also establishes development standards and build integrity through .gitignore, .pre-commit-config.yaml, MANIFEST.in, and a SECURITY.md policy.
(repo-wide) · high confidence
Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.
How this codebase got here
Score
- CAI 57 → 62 (+5.4)
- Rubric changed (rubric-2026.10.5 → rubric-2026.11.2) — scores are not directly comparable.
Lenses
- Code Health 55 → 55 (-0.1)
- Architecture 99 → 97 (-1.6)
- Maturity 46 → 63 (+17.0)
- Readiness 62 → 62 (+0.0)
- Security 82 → 75 (-7.2)
- Performance 100 → 100 (+0.0)
Resolved (8)
- Duplicate field/property definition. The config class lists 'mineru_parse_method' twice: once as a Field and once as a Property. This is likely a documentation error or a code generation artifact, but it presents as a confusing API surface where the same name maps to two different member types.
- Duplicate functionality across different types. 'RAGAnything.set_content_source_for_context' and 'document_context.set_document_source' appear to perform the same logical operation (setting the source for context extraction), but are exposed on different objects. This leads to API fragmentation.
- Inconsistent naming and scope for batch operations. 'process_documents_batch' and 'process_documents_with_rag_batch' have very similar signatures but unclear differentiation in naming (is the difference just the RAG integration step?). 'process_documents_batch_async' duplicates the sync version's intent but adds async. The distinction between 'batch' and 'batch with rag' is not immediately obvious from the name alone.
- Members sharing a duplicated core (4 members, 50+ identical tokens) (raganything/modalprocessors.py)
- Off-boarding risk: anonymized user #1
- Redundant methods with overlapping intent. Both methods appear to verify parser installation status. 'verify_parser_installation_once' suggests a cached or one-time check, while 'check_parser_installation' is generic. This creates confusion for the caller about which method to use for initialization checks.
- omml_extractor.extract_omml_equations (cognitive 26) (raganything/omml_extractor.py)
- utils.extract_section_path_from_content_list (cognitive 19) (raganything/utils.py)
New (14)
- EquationModalProcessor.generate_description_only (cognitive 17) (raganything/modalprocessors.py)
- FileTooLong: raganything/batch_parser.py (raganything/batch_parser.py)
- FileTooLong: raganything/utils.py (raganything/utils.py)
- Hotspot: raganything/raganything.py (raganything/raganything.py)
- Hotspot: raganything/utils.py (raganything/utils.py)
- Members sharing a duplicated core (4 members, 50+ identical tokens) (raganything/modalprocessors.py)
- Off-boarding risk: anonymized user #1
- Orphaned file (raganything/prompt_manager.py)
- Repeated repair: raganything/enhanced_markdown.py (raganything/enhanced_markdown.py)
- Repeated repair: raganything/modalprocessors_video.py (raganything/modalprocessors_video.py)
- Repeated repair: raganything/omml_extractor.py (raganything/omml_extractor.py)
- Repeated repair: raganything/resilience.py (raganything/resilience.py)
- Scattered collaborators
- utils._token_window_spans (cyclomatic 18) (raganything/utils.py)
Changes since last survey
- 27 commits — 8 feature/other, 19 fixes
By area
- (root) — 4 commits
- raganything/utils.py — 4 commits
- raganything/omml_extractor.py — 3 commits
- docs/batch_processing.md — 2 commits
- raganything/enhanced_markdown.py — 2 commits
- raganything/parser.py — 2 commits
- raganything/init.py — 1 commit
- raganything/batch_parser.py — 1 commit
- raganything/config.py — 1 commit
- raganything/mineru_content.py — 1 commit
- raganything/modalprocessors.py — 1 commit
- raganything/modalprocessors_audio.py — 1 commit
- raganything/modalprocessors_video.py — 1 commit
- raganything/parse_options.py — 1 commit
- raganything/query.py — 1 commit
- raganything/resilience.py — 1 commit
Notable commits
- fix: fix(batch): measure per-file timeouts from worker start (#434)
- fix: fix(batch): reparse files when their output directory is missing (#431)
- fix: fix(cache): invalidate parsed results when Docling options change (#432)
- fix: fix(charts): send rendered charts through the image pipeline (#452)
- fix: fix(config): read environment defaults at instantiation (#428)
- fix: fix(markdown): preserve relative images in WeasyPrint PDF output (#426)
- fix: fix(modalprocessors): don't name every regex-fallback entity unknown_entity (#416)
- fix: fix(omml): convert limits under a function name and Word's function names (#422)
- fix: fix(omml): leave tracked deletions out of extracted equations (#421)
- fix: fix(query): preserve complete retrieved image paths (#445)
- fix: fix(release): complete Linux setup hints and source distribution (#451)
- fix: fix(resilience): ignore outcomes from older circuit recovery cycles (#425)
- fix: fix(utils): keep every table cell in its column (#423)
- fix: fix(video): close MoviePy readers after audio extraction errors (#427)
- fix: fix: finish re-ingested docs, keep README equation formulas, restrict WeasyPrint resource loading (#446)
- fix: fix: pin MinerU to 3.x with a clear version check and MinerU's real error output (#449)
- fix: fix: pin av<19 for faster-whisper; document Docling install, parser selection and VLM streaming (#420)
- fix: fix: preserve Japanese kana span joining without removing word spaces (#441)
- fix: fix: skip null caption items (#444)
- change: Discard empty Whisper transcription segments (#438)
- …and 7 more
Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.
Survey your own repository
HKUDS/RAG-Anything was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.
About this page
- The score is its most recent published measurement, taken on 4 November 2026 at a pinned commit. It is not a live figure and does not change until the project is measured again.
- Measured at commit 2cd2b3329c470ccb53923f8a6d320596b34e56c5 — the exact code this score is about.
- Scored under rubric-2026.11.2 — the same rubric and the same method as every other entry in this index.
- Measured by watchdog.canine.dev using codehealth-analyzer measure-2026.11.2+preprod-f7b8aaa3a438.