Release Note V1.4.9.9

Important Notes

Add Workspace Isolation for Pipeline Status and In-memory Storage: Multiple LightRAG instances with distinct workspaces can be created simultaneously, marking a significant advancement toward seamless workspace switching within a single LightRAG server.
Add Workspace Vector Data Isolation by Model Name and Dimension for PostgreSQL and Qdrant: Previously, LightRAG used a single collection/table for difference embedding model and dimension, which caused dimension mismatch crashes or data pollution in multi-workspace.
Dimension Selection is Supported for OpenAI and Gemini Embedding model with new env var introduced: EMBEDDING_SEND_DIM
Add LLM Cache Migration and LLM Query Cache Cleanup Tools Between Different KV Storage: enabling users to switch storage backends without losing cached extraction and summary data.
Enhanced Enhanced DOCX Extraction with Table Content Support.
Enhanced XLSX extraction with proper handling of tab and newline characters within cells.
Fix Critical Security Vulnerability in React Server Components: #2494
Add Automatic Text Truncation Support for Embedding Functions: OpenAI embedding function now respect max_token_size value in EmbeddingFunc, and automatic truncate input text to prevent API errors caused by texts exceeding model token limits.

What's Breaking

Rename params of chunking function: If you incorporate the chunking function into LightRAG and pass parameters by name, corresponding code updates are required.

def chunking_by_token_size(
    tokenizer: Tokenizer,
    content: str,
    split_by_character: str | None = None,
    split_by_character_only: bool = False,
    chunk_overlap_token_size: int = 100,
    chunk_token_size: int = 1200,
) -> list[dict[str, Any]]:

Inject an embedding_func with model_name by wrap_embedding_func_with_attrs:

@wrap_embedding_func_with_attrs(
    embedding_dim=1536, max_token_size=8192, model_name="text-embedding-3-small"
)
@retry(
    stop=stop_after_attempt(3),
    wait=wait_exponential(multiplier=1, min=4, max=60),
    retry=(
        retry_if_exception_type(RateLimitError)
        | retry_if_exception_type(APIConnectionError)
        | retry_if_exception_type(APITimeoutError)
    ),
)
async def embedding_func(texts: list[str]) -> np.ndarray:
    client = AzureOpenAI(
        api_key=AZURE_OPENAI_API_KEY,
        api_version=AZURE_EMBEDDING_API_VERSION,
        azure_endpoint=AZURE_OPENAI_ENDPOINT,
    )
    embedding = client.embeddings.create(model=AZURE_EMBEDDING_DEPLOYMENT, input=texts)

    embeddings = [item.embedding for item in embedding.data]
    return np.array(embeddings)

rag = LightRAG(
      working_dir=WORKING_DIR,
      llm_model_func=llm_model_func,
      embedding_func=embedding_func,
)

To ensure seamless transition, legacy code injecting embedding_func without model_name will continue to interface with the original non-suffixed vector tables.

What's New

Feat: Add Chain of Thought Support for Gemini LLM by @danielaskdd in #2326
Feat: Add Optional Embedding Dimension Control with OpenAI API by @danielaskdd in #2328
Feat: Add Gemini Embedding Support to LightRAG by @danielaskdd in #2329
Feat: Add LLM Cache Migration Tool by @danielaskdd in #2330
Feat: Add LLM Query Cache Cleanup Tool by @danielaskdd in #2335
Support async chunking func to improve processing performance when a heavy chunking_func is passed in by user by @tongda in #2336
Add ollama cloud support by @LacombeLouis in #2348
Feat: Add Workspace Isolation for Pipeline Status and In-memory Storage by @danielaskdd in #2369
feat: add vchordrq vector index support for PostgreSQL by @wmsnp in #2378
Feat: Enhanced DOCX Extraction with Table Content Support by @danielaskdd in #2383
Feat: Enhance XLSX Extraction by Adding Separators and Escape Special Characters by @danielaskdd in #2386
Optimize for OpenAI Prompt Caching: Restructure entity extraction pro… by @Ghazi-raad in #2426
feat: Vector Storage Model Isolation with Automatic Migration by @BukeLy in #2391
feat: Implement Vector Database Model Isolation and Auto-Migration by @danielaskdd in #2513
feat: Add Automatic Text Truncation Support for Embedding Functions by @danielaskdd in #2523

What's Changed

Fix: Remove Duplicate Entity/Realtion Tracking Deletion in adelete_by_doc_id by @danielaskdd in #2322
Fix spelling errors in the "使用PostgreSQL存储" section of README-zh.md by @huangbhan in #2327
Add dimensions parameter support to openai_embed() by @yrangana in #2323
Fix Gemini driver retry mechanism by @danielaskdd in #2331
HotFix: Restore OpenAI Streaming Response & Refactor keyword_extraction Parameter by @danielaskdd in #2334
Refactor: Migrate PDF processing dependency from pypdf2 to actively pypdf by @danielaskdd in #2338
Fix: Prevent UnicodeEncodeError in JSON storage operations by @danielaskdd in #2344
Remove deprecated response_type parameter from query settings UI by @danielaskdd in #2345
Refactor: Optimize write_json for Memory Efficiency and Performance by @danielaskdd in #2346
Refact: Remove blocking dependency installation from document upload handlers by @danielaskdd in #2350
Refact: Implement Lazy Configuration Initialization for API Server by @danielaskdd in #2351
Refact: Enhance DOCLING integration with lazy loading and macOS safeguards by @danielaskdd in #2352
Fix: Robust error handling for async database operations in graph storage by @danielaskdd in #2356
Update the value corresponding to the extracted entity relationship keywords by @sleeepyin in #2358
Add macOS fork safety check for Gunicorn multi-worker mode by @danielaskdd in #2360
Refact: Add Embedding Token Limit Configuration and Improve Error Handling by @danielaskdd in #2359
Refact: Add Embedding Dimension Validation in EmbeddingFunc by @danielaskdd in #2368
test: Convert test_workspace_isolation.py to pytest style by @BukeLy in #2371
refactor(chunking): rename params and improve docstring for chunking by @EightyOliveira in #2379
Fix: Add chunk token limit validation with detailed error reporting by @danielaskdd in #2389
Fix: Remove redundant exception logging to eliminate pytest shutdown errors by @danielaskdd in #2390
issue-2394: use deployment variable instead of model for embeddings API call by @Amrit75 in #2395
Refactor: Centralize keyword_extraction parameter handling in OpenAI LLM implementations by @danielaskdd in #2401
Refact: Consolidate Azure OpenAI and OpenAI implementations by @danielaskdd in #2403
Update README.md by @chaohuang-ai in #2408
Update README.md by @chaohuang-ai in #2409
feat: create copilot-setup-steps.yml by @netbrah in #2410
Fix: Add Comprehensive Retry Mechanism for Neo4j Storage Operations by @danielaskdd in #2417
Refact: Allow API Server to Start Without Built WebUI Assets by @danielaskdd in #2418
fix:exception handling order error by @EightyOliveira in #2421
Doc: Update README examples to prevent double-wrapping of embedding functions by @danielaskdd in #2432
Fix: Add configurable model support for Jina embedding by @danielaskdd in #2433
Fix typos discovered by codespell by @cclauss in #2434
Update README.md by @chaohuang-ai in #2439
Fix KaTeX chemistry formula rendering (\ce command) not working by @danielaskdd in #2443
fix(postgres): Add CASCADE to AGE extension creation for automatic dependency resolution by @danielaskdd in #2446
Add Python 3.13 and 3.14 to the testing by @cclauss in #2436
Keep GitHub Actions up to date with GitHub's Dependabot by @cclauss in #2435
chore: optimize Dependabot configuration with dependency grouping and PR limits by @danielaskdd in #2447
fix: Return existing track_id for duplicate documents in document insertion endpoints by @danielaskdd in #2469
feat: Add Cohere reranker config, chunking, and tests by @netbrah in #2411
update to react 19.2.1 by @mccahill in #2494
Fix links in README.md and README-zh.md by @johuellm in #2480
fix(prompt): use language parameter in keywords_extraction prompt by @jcfneto in #2491
feat: Update full-text index name to include workspace label for bett… by @netbrah in #2406
HotFix: MongoDB __post_init__ not called and refactor embedding validation by @danielaskdd in #2521
Fix nested EmbeddingFunc wrapping and improve embedding function safety by @danielaskdd in #2522
fix(postgres): Handle dollar-sign sequences in Cypher queries to prevent syntax errors by @danielaskdd in #2524

New Contributors

@huangbhan made their first contribution in #2327
@LacombeLouis made their first contribution in #2348
@sleeepyin made their first contribution in #2358
@wmsnp made their first contribution in #2378
@EightyOliveira made their first contribution in #2379
@Amrit75 made their first contribution in #2395
@chaohuang-ai made their first contribution in #2408
@netbrah made their first contribution in #2410
@cclauss made their first contribution in #2434
@mccahill made their first contribution in #2494
@johuellm made their first contribution in #2480
@Ghazi-raad made their first contribution in #2426
@jcfneto made their first contribution in #2491

Full Changelog: v1.4.9.8...v1.4.9.9

HKUDS/LightRAG v1.4.9.9 on GitHub

Release Note V1.4.9.9

Important Notes

What's Breaking

What's New

What's Changed

New Contributors

HKUDS/LightRAG v1.4.9.9
on GitHub