LlamaIndex • Aug 22, 2024

Introducing LlamaIndex 0.11

Python

LlamaIndex is delighted to announce that we have released the latest and greatest version of LlamaIndex for Python, version 0.11! There's been lots of updates since 0.10 was released, but here are a few highlights:

Workflows

We’ve introduced Workflows, an event-driven architecture for building complex gen AI applications. This replaces our previous experiment in this area, Query Pipelines, which should now be considered deprecated. We’re very excited about Workflows, so be sure to check out our new tutorial on how they work!

Instrumentation

We significantly upped our observability game with our new instrumentation feature, allowing better monitoring and debugging of LlamaIndex applications.

Property Graph Index

Users have been delighted with our significantly improved support for property graphs with our Property Graph Index.

Reduced package size

Everyone will be pleased to hear that we've substantially reduced the size of the llama-index-core package -- by 42%! We did this by removing OpenAI as a core dependency, adjusting how we depend on nltk, and making Pandas an optional dependency.

Other improvements and additions

Async streaming support for query engines and throughout the framework
A Structured Planning Agent to improve our agentic capabilities
Function Calling LLM to better handle tool-calling in LLMs
Chat Summary Memory buffer to improve our ability to maintain context in conversations

Plus lots more!

We've added hundreds of new features and bug fixes since 0.10, way too many to list!

Breaking changes

In addition to all those goodies, 0.11 does bring some breaking changes:

Pydantic V2

After a huge effort by Andrei we are able to move fully to Pydantic V2. If you were previously needing to use pydantic.v1 imports to get around our lack of support for v2, you can now remove those.

Pydantic V2 means LlamaIndex Pydantic types will work directly with FastAPI and other frameworks that use V2.

This was a big change, so if you see any bugs please report them!

ServiceContext removed

Deprecated since 0.10, the ServiceContext object is now completely removed in favor of using the Settings object. If you import and use ServiceContext in 0.11, you will get an error pointing you to our migration docs.

LLMPredictor removed

Also related to Settings is the removal of the LLMPredictor. You should instead use the LLM class, which is a drop-in replacement. Using LLMPredictor in 0.11 will raise an error.