v.0.0.3

What's Changed

• lint in #22
• Add JSON return format support, add PyPI package for API connect, return all citations instead of the winning one only in #25
• Improve exception handle

Full Changelog: v.0.0.2...v.0.0.3

v.0.0.2

Highlight

• Fully integrated with DSPy with MIPROv2 optimizer
• Changed default embedding and rerank inference to Infinity API server, faster and more stable
• Fixed CloudFlare 524 timeout error
• Changed search backend to https://search.ittia.net
• Added examples on how to create wiki_dpr index and start a retriever server

What's Changed

• Add DSPy pipeline in #7
• add api key support to LlamaIndex OllamaEmbedding in #9
• change all LLM calling to DSPy, increase citation token limit in #10
• change base image to CUDA, change to dspy.Retrieve in #11
• add script to create dataset based on HotPotQA, update infra in #12
• add wiki_dpr retriever for DSPy compile in #15
• change to multi-sources mode in #17
• change API response to stream in #18
• move pipelines to one single class, change to streaming search backend in #19
• change default search backend, update endpoint in #20

v.0.0.1

Features:

• For each statement, generate multiple verdicts against every web search page.
• Weight verdicts into one final and show related context and sources.
• Retrieval: LlamaIndex auto merging retriever, embedding and rerank.

What's next:

• More sophisticated pipeline using DSPy etc.
• Implement LLM features: multi-shot, chain of thought, etc.