github HKUDS/LightRAG v1.5.8rc2

one hour ago

What's Changed

  • fix(jina): stop retrying permanent 4xx errors by @Shizoqua in #4028
  • fix(pipeline): round-trip sidecar URIs for Windows drive paths by @rtmalikian in #4273
  • fix(rerank): stop retrying permanent errors in generic_rerank_api by @danielaskdd in #4420
  • fix(query): return the failure shape from aquery_data when no keywords are extracted by @Zhuoxi2000 in #4419
  • fix(postgres): give AGE a valid graph name for long or digit-leading workspaces by @zzjiho in #4408
  • fix(llm): retry malformed OpenAI completion responses by @ZXLyyds100 in #4279
  • fix(gemini): convert safety_settings mapping to SDK list shape by @Shizoqua in #4423
  • fix(nano): stop delete() from swallowing queue failures by @Shizoqua in #4424
  • fix(llm): close Zhipu clients after chat and embedding calls by @remote-controlled-man in #4427
  • test(rerank): cover ali_rerank aliyun request and response format by @Shizoqua in #4425
  • fix(mongo): propagate KV deletion failures to document recovery by @dajiaohuang in #4167
  • fix(postgres): propagate KV and doc-status delete failures by @danielaskdd in #4432
  • fix(postgres): escape LIKE wildcards in PGGraphStorage.search_labels by @MohammadHijjawi97 in #4431
  • fix: honor disabled DOCX table header markers by @LimbC-C in #4430
  • fix(opensearch): reject incomplete graph and status deletes by @dajiaohuang in #4169
  • fix(docx): honor OOXML on/off values everywhere and repeat only leading header rows by @danielaskdd in #4433
  • fix(openai): do not close a caller-supplied http_client by @GautamVhavle in #4237
  • fix(api): delete the archived source of a name ending in _NNN by @MohammadHijjawi97 in #4428
  • fix(api): preserve temporary-prefixed sources for deferred parsing by @dajiaohuang in #4173
  • fix(neo4j): truncate an oversized APOC subgraph from the walk already run by @wkpark in #4216
  • fix(hf): pass resolved max_length to truncate hf_embed inputs by @Shizoqua in #4138
  • fix(parser): 修复 UTF-8 BOM 污染正文并绕过空文件校验 by @2gg-bit in #4132
  • fix(rerank): 修复重排分块破坏 Unicode 字符 by @2gg-bit in #4131
  • fix(parser): accept UTF-16/UTF-32 text files with a BOM by @danielaskdd in #4434
  • fix(ollama): consume both legacy extraction flags by @xyi74976-del in #4289
  • fix(vlm): accept case-insensitive image data URLs by @Elon-Jiang in #4287
  • fix(openai): preserve client-configured default headers by @xujiantop-crypto in #4152
  • fix(nvidia): 修复 base64 嵌入响应无法返回数值向量 by @2gg-bit in #4135
  • fix: enforce the total token budget for KG query context by @dafyy321-pixel in #4221
  • fix(llama_index): honor embedding batching and cache policies by @lux-liang in #4200
  • fix(llama_index): handle the Settings singleton without set_global by @heranliu in #4209
  • fix(config): register provider options for the effective CLI binding by @dajiaohuang in #4171
  • docs(llm): document llama_index_complete's settings argument as ignored by @danielaskdd in #4436
  • ci: split the offline suite into four parallel directory shards by @danielaskdd in #4437
  • fix(tools): open the repair plan for writing before os.fsync by @simpleqt in #4251
  • fix: forward configured dimensions to Bedrock Titan V2 by @LimbC-C in #4284
  • fix(parser): preserve Unicode offsets when slicing HTML tables by @danielaskdd in #4438
  • fix(parser): keep w:cr, w:ptab and w:noBreakHyphen in native DOCX text by @L4XB in #4239
  • fix(parser): read w:cr, w:ptab and w:noBreakHyphen in smart-heading features by @danielaskdd in #4439
  • fix(parser): list only ancestor headings in MinerU/Docling parent_headings by @hardness1020 in #4227
  • fix(parser): preserve OMML function names across runs and scripts by @shuaiyafei in #4259
  • Fix parser bundles rejecting dot-prefixed filenames by @IlyaKhalafi in #4219
  • fix(parser): resolve dot-prefixed MinerU image refs inside the bundle by @danielaskdd in #4440
  • fix(webui): ignore superseded asynchronous option requests by @dajiaohuang in #4181
  • fix(sidecar): preserve dot-prefixed asset names by @Akaliup in #4291
  • fix(webui): preserve composing text when confirming IME candidates by @dajiaohuang in #4177
  • fix(examples): forward the Gemini embedding dimension by @dajiaohuang in #4185
  • fix(hf): preserve prompts and history during chat template fallback by @danielaskdd in #4441
  • fix(examples): create the configured demo log directory by @dajiaohuang in #4189
  • fix(examples): keep initialization errors visible during cleanup by @dajiaohuang in #4191
  • fix(reproduce): keep repeated query output a single JSON array by @dajiaohuang in #4194
  • fix(setup): keep retry diagnostics out of captured configuration by @dajiaohuang in #4183
  • fix(reproduce): retain the intended trailing summary window by @dajiaohuang in #4192
  • test: pin utf-8 and posix paths in four source-reading tests by @qhxwork in #4269
  • fix(parser): 保留行内代码和转义后的 Markdown 图片示例 by @2gg-bit in #4133
  • fix(tests): stop the offline suite from pip-installing provider SDKs by @UltimateWilliamWu in #4275
  • perf(docx): skip numbering regexes for body text (CB1 gate 7.7% faster) by @dexhunter in #4298
  • fix(api): store login rate-limit keys as fixed-size digests by @danielaskdd in #4442
  • fix(chunker): recover F-token _source_span at emoji boundaries by @leilei3167 in #4116
  • test: gate POSIX-only assertions so the offline suite runs on Windows by @qhxwork in #4270
  • fix: decode Unicode surrogate pairs in streamed text by @LimbC-C in #4274
  • test(config): use symbolic errno values in bind lock tests by @danielaskdd in #4453
  • fix(bedrock): retry on dropped connection errors by @Shizoqua in #4447
  • fix(docx): keep a table cell holding "" from ending the table tag by @Dev-next-gen in #4450
  • test(rerank): cover jina_rerank request and response format by @Shizoqua in #4445
  • test: remove test smells from 18 test modules (no production code changes) by @pedrorivald in #4290
  • fix(qdrant): run vector upsert flush off the event loop by @Shizoqua in #4446
  • fix(lollms): preserve text during streamed generation by @remote-controlled-man in #4281
  • fix(table): keep a "" cell from ending any re-emitted table tag by @danielaskdd in #4454
  • fix: validate custom chunk token limits before journaling by @adity982 in #3930
  • fix(evaluation): tokenize CJK text in the offline retrieval check by @UltimateWilliamWu in #4276
  • fix(lmdeploy): retry on transient engine errors by @Shizoqua in #4458
  • test(utils): cover generate_reference_list_from_chunks ordering by @Shizoqua in #4457
  • fix(networkx): keep bounded iteration live and name the mutation cause by @Shizoqua in #4459
  • docs: correct parser, multimodal, and setup guidance by @dajiaohuang in #4242
  • docs: fix some typos in API server docs and guides. by @Superb0431 in #4085
  • fix(chunker): verify token-window positions before attaching provenance by @xixi-box in #4463
  • fix(kg): retry transient Apache AGE label-cache dangling pointer errors in PGGraphStorage by @chengwudi1 in #4212
  • fix(kg): strip lone surrogates from NanoVectorDB records at upsert by @rtmalikian in #4472
  • fix(parser): count HTML table rows with omitted end tags by @shuaiyafei in #4260
  • fix(core): validate every custom KG file_path before writing by @danielaskdd in #4473
  • fix(bedrock): honor LLM_TIMEOUT in bedrock_complete for library callers by @kimnamu in #4076
  • fix(kg): refuse a scheduling cursor that is not a two-element list by @simpleqt in #4253
  • fix(milvus): bound the collection-load wait (default 600s), with a typed load error by @AdalbertoDam in #4230
  • fix(vector): clear every embedding model's vectors on workspace clear by @inamdarmihir in #4263
  • build(deps): bump the llm-providers group with 3 updates by @dependabot[bot] in #4466
  • build(deps): bump pymongo from 4.18.3 to 4.19.0 in the storage group by @dependabot[bot] in #4467
  • build(deps): bump the data-processing group with 2 updates by @dependabot[bot] in #4468
  • build(deps): bump the dev-tools group with 2 updates by @dependabot[bot] in #4470
  • fix: clean up incomplete uploads after cancellation by @dafyy321-pixel in #4223
  • build(deps): bump the web-framework group with 2 updates by @dependabot[bot] in #4469
  • fix(llm): honor LLM_TIMEOUT in every provider client, incl. per-role and 0 by @danielaskdd in #4474
  • fix(llm): treat a non-positive LLM_TIMEOUT as no timeout by @Lesereingrape in #4098
  • fix(query): report a failed query as an error instead of "no context" by @danielaskdd in #4478
  • fix(embedding,rerank): hand EMBEDDING_TIMEOUT / RERANK_TIMEOUT to the clients by @danielaskdd in #4479
  • fix(parser): preserve literal Markdown marker examples by @dakjdakd in #4204
  • fix(webui): move ThemeProviderContext out of the component file by @ankit2235 in #4127
  • fix(webui): stop lightrag-stream test leaking a fake axios into later files by @ankit2235 in #4126
  • fix(chunker): stop F-token windows once one reaches the end by @Yuilona in #4199
  • fix(reproduce): align judge answers by their saved query by @dajiaohuang in #4187
  • fix(migration): use authoritative chunk tracking for legacy anchors by @dajiaohuang in #4157
  • test(pg): smoke-test the legacy anchor migration on PostgreSQL by @danielaskdd in #4481
  • fix(auth): reject passwords over bcrypt's 72-byte limit in hash_password by @MohammadHijjawi97 in #4489
  • fix(archive): skip occupied timestamp filenames by @passionworkeer in #4490
  • fix(rerank): reject negative top_n in generic_rerank_api by @Shizoqua in #4487
  • fix(gemini): reraise the SDK error once retries are exhausted by @MohammadHijjawi97 in #4482
  • fix(query): never return or cache a reasoning-only answer by @cvidaillac in #4483
  • fix(chunker): reject chunk sizes below a configuration minimum by @Shizoqua in #4488
  • fix: document streaming queries and annotate query() stream result as AsyncIterator by @wToTw123 in #4480
  • fix(extraction): keep entities that share a line with other records by @SulimanAbdulrazzaq in #4099
  • fix(bedrock): join all text blocks instead of only the first by @Shizoqua in #4060
  • fix(rebuild): attribute shared extraction cache rows via llm_cache_list by @danielaskdd in #4494
  • feat(rebuild): re-extract surviving chunks missing their initial extraction by @danielaskdd in #4495
  • fix(docx): treat a gridSpan below 1 or without an integer value as one column by @RizgarOzan in #3962
  • fix(parser): make filename-hint validation and classification agree by @RizgarOzan in #3995
  • feat(tools): add an offline GC for orphaned LLM cache rows by @cgycorey in #4155
  • fix(memgraph): bind workspace label as a parameter in string-literal context by @YashvantHange in #4119
  • fix(parser): deny the IPv4-translated SIIT prefix in the image-download guard by @danielaskdd in #4500
  • fix(chunker): supported sidecar resolver and lazy ChunkingContext (follow-up to #4010) by @mikelju in #4136
  • fix(neo4j): route Korean and kana label searches to the CJK query path by @hyojin52 in #4128
  • feat(webui): transition WebUI to account-only authentication (#4125) by @Pioneer-Weirdo in #4121
  • fix(query): recover keyword JSON followed by bracketed prose by @MohammadHijjawi97 in #4118
  • fix(lollms): make the retry predicate match what aiohttp raises by @Pioneer-Weirdo in #4122
  • build(deps): bump lucide-react from 1.70.0 to 1.72.0 in /lightrag_webui in the ui-components group by @dependabot[bot] in #4501
  • build(deps-dev): bump the build-tools group in /lightrag_webui with 4 updates by @dependabot[bot] in #4502
  • build(deps): bump the frontend-minor-patch group in /lightrag_webui with 3 updates by @dependabot[bot] in #4503
  • fix(pipeline): refuse drive-relative Windows sidecar URIs on every Python by @danielaskdd in #4512
  • test(faiss): make fingerprint-fence tests independent of mtime resolution by @danielaskdd in #4514
  • test(parser): handle missing native Cairo in Markdown tests by @leilei3167 in #4513
  • docs(readme): explain embedding timeouts and EMBEDDING_TIMEOUT by @danielaskdd in #4516
  • Make JSON mode optional for query keyword extraction by @chengwudi1 in #4026
  • fix(kg): fall back when env-backed storage limits are unusable by @Lesereingrape in #4097
  • fix(utils): warn on unrecognized boolean env values instead of reading False by @danielaskdd in #4518
  • fix(tools): resolve audit provenance from chunk tracking, not the capped source_id by @JessYanCoding in #4084
  • fix(postgres): scale hnsw.ef_search with vector query top_k by @Shizoqua in #4081
  • fix(custom-kg): merge provenance into existing entities and relations by @danielaskdd in #4506
  • fix(pipeline): fail a commit whose chunk-tracking write was retained by @danielaskdd in #4521
  • fix(mongo): scale vector search numCandidates with top_k by @Shizoqua in #4079
  • fix(networkx): snapshot once for batch graph reads by @Shizoqua in #4080
  • fix(ollama): retry real connection error type by @Shizoqua in #4078
  • fix(lollms): reject unsuccessful HTTP responses by @danielaskdd in #4522
  • fix(lollms): reject unsuccessful HTTP responses by @yeliheng-010 in #4057
  • fix: avoid mutating caller chunker config by @remote-controlled-man in #4073
  • fix(gemini): parse GOOGLE_GENAI_USE_VERTEXAI like the google-genai SDK by @huiq777 in #4074
  • fix(evaluation): dedupe expected docs in recall_at denominator by @Shizoqua in #4061
  • fix: don't swallow a cancelled document deletion in adelete_by_doc_id by @YHC66 in #4052
  • fix(anthropic): extract answer text from the correct content block by @Shizoqua in #3973
  • fix(milvus): always use the COSINE metric and re-index legacy L2/IP collections by @danielaskdd in #4527
  • fix(parser): apply the forbidden-character rule on the engine-param dict path by @RizgarOzan in #4005
  • fix(chunker): keep token-sized segments through no-op separators by @RaycarlLei in #3979
  • ci: cancel superseded pull request runs on a new push by @danielaskdd in #4530
  • test(llm): cover TokenTracker propagation through role wrappers by @lorenzozanee in #3969
  • fix: add storage consistency checks and bounded rebuilds by @leninworld in #4071
  • fix(rerank): sort results by score before use by @Shizoqua in #3961
  • fix(multimodal): anchor table context on the table not a citation by @Shizoqua in #3965
  • fix(kg): stop edge weight double counting within one batch by @Shizoqua in #3964
  • fix(kg): favor established entity type on a tie by @Shizoqua in #3959
  • fix(parser): pad Markdown pipe-table rows to a rectangular grid by @nurlanmnn in #4493
  • fix(examples): migrate the query generator off GPT-4o mini by @JunaidAhamed-7777 in #4499
  • fix(extraction): fail the merge on an empty description summary by @zzjiho in #4498
  • test(pipeline): pin the cause-cycle bound and merge duplicate flush tests by @Shizoqua in #4509
  • fix(extraction): cap file paths when rebuilding an entity from its edges by @Shizoqua in #4510
  • fix(export): export each undirected relation once by @67-luck in #4515
  • fix(neo4j): avoid full-text index name collisions across workspaces by @Shizoqua in #4486
  • fix(mongo): add tie-breaker to doc pagination sort by @Shizoqua in #4461
  • fix(query): skip duplicate rerank indices before top-k by @wToTw123 in #4523
  • fix(markdown): do not open a fence on a one-line backtick code span by @MohammadHijjawi97 in #4528
  • fix(query): report keywords that contain a comma whole in query metadata by @Zhuoxi2000 in #4529
  • fix(query): validate kg_chunk_pick_method and fall back to WEIGHT when VECTOR is unavailable by @SDLSY in #4531
  • fix(markdown): treat $$...$$ followed by prose as text by @chuenchen309 in #4526
  • fix(llm): handle nullable counts in TokenTracker by @wToTw123 in #4524
  • fix(jina): align embedding rows with input indices by @Ymmcc in #4532
  • test(retry): pin the aiohttp retry table once, one wiring case per branch by @danielaskdd in #4533
  • fix(nvidia): reraise original exception after exhausted retries by @aloneangel45 in #4536
  • fix(voyageai): reraise original error instead of tenacity.RetryError by @Shizoqua in #4539
  • test(chunker): cover min_chunk_size merge-forward behavior in V chunker by @Shizoqua in #4543
  • fix(zhipu): drop stream kwarg before the SDK call to avoid a crash by @Shizoqua in #4541
  • fix(webui): turn off Tailwind automatic source detection by @danielaskdd in #4546
  • feat(webui): load Prism languages and KaTeX extensions on demand by @danielaskdd in #4547

New Contributors

Full Changelog: v1.5.8rc1...v1.5.8rc2

Don't miss a new LightRAG release

NewReleases is sending notifications on new releases.