RetrievalQAWithSourcesChain

chain_type is important

run

1. call()

2. get_docs() get relevant documents under max token length https://github.com/langchain-ai/langchain/blob/master/libs/langchain/langchain/chains/qa_with_sources/retrieval.py
variable initialization and mapping function in https://github.com/langchain-ai/langchain/blob/master/libs/langchain/langchain/chains/qa_with_sources/loading.py
• question 으로만 source 찾는 한계
• hf native보다 너무 느림 because of multi inference in combine strategy
• input_documents are mapped in BaseCombineDocumentsChain.combine_docs()

3. combine_docs() 함수로 combine
• stuff - Stuff all documents into one prompt and pass to LLM
https://github.com/langchain-ai/langchain/blob/master/libs/langchain/langchain/chains/combine_documents/stuff.py
• refine -
• map_reduct - use stuff, mapreduce, reduce all so complicate implementation
• https://github.com/langchain-ai/langchain/blob/master/libs/langchain/langchain/chains/combine_documents/reduce.py
• https://github.com/langchain-ai/langchain/blob/master/libs/langchain/langchain/chains/combine_documents/map_rerank.py
• map_rerank
• https://github.com/langchain-ai/langchain/blob/master/libs/langchain/langchain/chains/combine_documents/map_rerank.py

4. https://github.com/langchain-ai/langchain/blob/master/libs/langchain/langchain/chains/llm.py
• 에서 infrerence with doc
1. predict()
2. _call()
3. generate()
1. prepare prompts

5. https://github.com/langchain-ai/langchain/blob/master/libs/langchain/langchain/llms/base.py
1. generate_prompt
2. generate()