Eric Zhu

One-Permutation hashing seems to speed up MinHash creation without loosing much accuracy. Related papers: [Lazo](https://ieeexplore.ieee.org/document/8731486), [FLASH](https://dl.acm.org/doi/10.1145/3183713.3196925). We can try this out. However this really depends on the accuracy-speed trade off....

This is going to be an on-going thread about reducing memory usage of LSH by using reservoir sampling and sharing for hash buckets, as proposed by [Wang et al.](https://dl.acm.org/doi/10.1145/3183713.3196925). Related...

To address some of the issues raised in #4

adding command line tool for dumping all metadata

11

This address issue #98

Cross-document queries (examples)

1

Some examples of cross-document queries -- queries that must be answered using multiple documents. For now I am testing two documents.

Evaluate text-2-sql capability using Spider benchmark

7

We want to benchmark LlamaIndex's performance for complex queries on multiple domains, and measure how each iteration of LLM improves its Text-to-SQL capability, thus this PR.

Improve text-to-sql performance with updated prompt and stop token

1

1. On spider benchmark dev set, execution accuracy increased to 70% from 50% (on a 1% sample queries). 2. Added `stop_token` to `Prompt` class to allow new prompts to specify...

AzureChatOpenAI for Azure Open AI's ChatGPT API

1

Add support for Azure OpenAI's ChatGPT API, which uses ChatML markups to format messages instead of objects. Related issues: #1591, #1659

https://github.com/hwchase17/langchain/blob/master/langchain/agents/agent_toolkits/sql/toolkit.py#L38 SqlDatabaseToolkit should have custom llm_chain field for QueryCheckerTool. This is causing issues when OpenAI is not available, as the QueryCheckerTool will automatically use OpenAI.

Add human as a tool

1

Human can help AI. #1871