Jaankoppe-llama-index

New Features
- add non-linear embedding adapter (7658)
- Add "finetune + RAG" evaluation to knowledge fine-tuning notebook (7643)

Bug Fixes / Nits
- Fixed chunk-overlap for sentence splitter (7590)

New Features
- Added `AGENT_STEP` callback event type (7652)

Bug Fixes / Nits
- Allowed `simple` mode to work with `as_chat_engine()` (7637)
- Fixed index error in azure streaming (7646)
- Removed `pdb` from llama-cpp (7651)

New Features
- guide: fine-tuning to memorize knowledge (7626)
- added ability to customize prompt template for eval modules (7626)

Bug Fixes
- Properly detect `llama-cpp-python` version for loading the default GGML or GGUF `llama2-chat-13b` model (7616)
- Pass in `summary_template` properly with `RetrieverQueryEngine.from_args()` (7621)
- Fix span types in wandb callback (7631)

Bug Fixes
- Make sure context and system prompt is included in prompt for first chat for llama2 (7597)
- Avoid negative chunk size error in refine process (7607)
- Fix relationships for small documents in hierarchical node parser (7611)
- Update Anyscale Endpoints integration with full streaming and async support (7602)
- Better support of passing credentials as LLM constructor args in `OpenAI`, `AzureOpenAI`, and `Anyscale` (7602)

Breaking Changes
- Update milvus vector store to support filters and dynamic schemas (7286)
- See the [updated notebook](https://gpt-index.readthedocs.io/en/stable/examples/vector_stores/MilvusIndexDemo.html) for usage
- Added NLTK to core dependencies to support the default sentence splitter (7606)

New Features
- Added support for ElasticSearch Vector Store (7543)

Bug Fixes / Nits
- Fixed small `_index` bug in `ElasticSearchReader` (7570)
- Fixed bug with prompt helper settings in global service contexts (7576)
- Remove newlines from openai embeddings again (7588)
- Fixed small bug with setting `query_wrapper_prompt` in the service context (7585)

Breaking/Deprecated API Changes
- Clean up vector store interface to use `BaseNode` instead of `NodeWithEmbedding`
- For majority of users, this is a no-op change
- For users directly operating with the `VectorStore` abstraction and manually constructing `NodeWithEmbedding` objects, this is a minor breaking change. Use `TextNode` with `embedding` set directly, instead of `NodeWithEmbedding`.

New Features
- add embedding adapter fine-tuning engine + guide (7565)
- Added support for Azure Cognitive Search vector store (7469)
- Support delete in supabase (6951)
- Added support for Espilla vector store (7539)
- Added support for AnyScale LLM (7497)

Bug Fixes / Nits
- Default to user-configurable top-k in `VectorIndexAutoRetriever` (7556)
- Catch validation errors for structured responses (7523)
- Fix streaming refine template (7561)