Release 2.16.0

2.16.0

New Features

• Update documents for unstructured indexes (#1030). The update_documents endpoint is now supported for unstructured indexes created after version 2.16. This allows you to update documents by modifying existing non-tensor fields or adding new fields without re-indexing the entire document. Check here for more details.
• Configurable max Vespa disk utility (#1124). Users can set the VESPA_DISK_USAGE_LIMIT environment variable (ranging from 0 to 1) to adjust the disk usage limit for Vespa when using Marqo.

Bug Fixes and Minor Changes

• Fix a bug where Marqo can not load the OpenCLIP model tokenizer when the model name has a hf-hub: prefix (#1126).
• Clean up the logs when starting Marqo (#1137).

Contributor Shout-Outs

• A huge thanks to our 4.8k stargazers for your continued support!
• Thanks a lot for the discussion and suggestions in our community. Join us on Slack and our forum today!

Release 2.15.2

2.15.2

Bug fixes and minor changes

• Fix an issue that Marqo can not load legacy OpenAI CLIP models (#1116).

• Improve the performance for large-limit searches (#1116).

• Fix hybrid search behaviour where any error is raised as a generic 500 error (#1116). Users will now receive the correct error code and error messages when one or multiple errors occur in a hybrid search with the disjunction retrieval method.

Release 2.13.6

2.13.6

Bug fixes and minor changes

• Improved error messages for hybrid search (#1108). Errors that occur in the tensor or lexical search portion of hybrid search are now correctly raised instead of becoming generic 500 errors.

• Add missing support for legacy OpenAI CLIP models (#1107). This fix allows old OpenAI CLIP models (eg. ViT-B/32) to be loaded properly.

• Bug fix for searches using CustomVectorQuery. Searches of this query type will no longer fail due to result not being JSON-serializable (#1109).

Release 2.15.1

2.15.1

Bug fixes and minor changes

• Fix a bug where numeric map fields are not returned when searching with attributes_to_retrieve parameter for unstructured indexes created prior to Marqo 2.13 (#1097).

• Fix a bug where numeric fields, numeric map fields, boolean fields and string array fields are not returned when searching with attributes_to_retrieve parameter for unstructured indexes created with Marqo 2.13 or later (#1097).

• Fix a bug where document-processing element is removed from the services.xml config file when bootstrapping the vector store (#1097).

Release 2.13.5

2.13.5

Bug fixes and minor changes

• Improve performance of processing large search responses (#1091). Optimizations include: removing unnecessary validation, utilizing orjson for faster JSON serialization, and optimizing post-search field selection (for unstructured indexes).

Release 2.15.0

2.15.0

New Features

• Global Score Modifiers for Hybrid Search (#1082). Introduce global score modifiers for hybrid search, allowing fine-tuned adjustments to RRF scores in combined result lists. This enhancement provides better control over returned results in hybrid search scenarios. Use the rerankDepth parameter to control the number of hits to rerank. For detailed usage, check here.

• Custom LanguageBind Model (#1072). Marqo now supports loading custom LanguageBind models from S3 buckets, URLs, or HuggingFace model cards. Fine-tune your own LanguageBind model and integrate it with Marqo to achieve better in-domain results. For more details, see here.

• Add Marqtune models to the model registry (#1063). Marqtune models have been added to the model registry, with some models renamed to align with the Marqtune naming convention. This update improves consistency and makes it easier to identify models Marqo and Marqtune. The changes are fully backwards compatible.

Bug Fixes and Minor Changes

• Fix a bug where Marqo incorrectly inferred the modality of a field, even when the field is not a tensor field in an unstructured index (#1086).
• Resolve an issue where Languagebind models could only handle a single video or audio in a weighted search query. You can now provide multiple videos and audios in a weighted query seamlessly.(#1072).
• Improve memory usage when indexing image documents with LanguageBind models, enabling more efficient handling of image data. (#1072).

Contributor Shout-Outs

• A huge thanks to our 4.7k stargazers for your continued support!
• Thanks a lot for the discussion and suggestions in our community. Join us on Slack and our forum today!

Release 2.14.1

2.14.1

New features

• Add support for hf_transfer to accelerate the downloading speed of HuggingFace models by 10 to 30 times. See here for details about how to enable it (#1066).

• Add /healthz endpoint for Marqo container liveness checks, which performs a status check for CUDA devices and returns a 500 error if any existing CUDA devices become unavailable or run out of memory (#1068).

Bug fixes and minor changes

• Fix a bug where numeric map fields are not returned when searching with attributes_to_retrieve parameter for unstructured indexes created prior to Marqo 2.13 (#1062).

• Fix a bug where numeric fields, numeric map fields, boolean fields and string array fields are not returned when searching with attributes_to_retrieve parameter for unstructured indexes created with Marqo 2.13 or later (#1062).

• Fix a bug where document-processing element is removed from the services.xml config file when bootstrapping the vector store (#1075).

Release 2.13.4

2.13.4

Bug fixes and minor changes

• Fix a bug where document-processing element is removed from the services.xml config file when bootstrapping the vector store (#1075).

Release 2.13.3

2.13.3

Bug fixes and minor changes

• Fix a bug where numeric map fields are not returned when searching with attributes_to_retrieve parameter for unstructured indexes created prior to Marqo 2.13 (#1062).

• Fix a bug where numeric fields, numeric map fields, boolean fields and string array fields are not returned when searching with attributes_to_retrieve parameter for unstructured indexes created with Marqo 2.13 or later (#1062).

Release 2.14.0

2.14.0

New features

• FFmpeg-CUDA Support (#1030). Add GPU acceleration for video decoding by integrating FFmpeg with CUDA support. This feature significantly improves video processing performance, making video handling up to 5 times faster. Check here for guidance and requirements.

• Video and audio file size limits (#1012). Introduce configurable size limits for video and audio files in the add_documents, search, and embed endpoints. This enhancement allows users to manage and optimize resource usage effectively, ensuring smoother processing of multimedia content. Check here for more details.

• Upgrade to Python 3.9 (#1006). Upgrade the Marqo Docker image to use Python 3.9. With Python 3.8 reaching its End of Life (EOL), we have upgraded our platform to Python 3.9 to maintain security, compatibility, and access to ongoing support.

Bug fixes and minor changes

• Move NLTK resource downloads to Marqo's startup process and remove the unsafe punkt package. This avoids potential cold start issues and enhances security (#1040).

• Fix model serialization in OpenAPI specifications. This community-contributed PR resolves issues with OpenAPI spec generation and SwaggerUI by fixing incorrect type hints in the API definition, ensuring accurate model serialization and improving API documentation accessibility (#986).

• Add brief description for each endpoint in the OpenAPI specifications. This improves API documentation clarity for users (#1042).

Contributor shout-out

• Shoutouts to our valuable 4.7k stargazers!
• Thanks a lot for the heated discussion and suggestions in our community. We love to hear your thoughts and requests. Join our Slack channel and forum now.
• Special thanks to community contributor @gabauer for their impactful PR, helping improve Marqo for everyone!