Do we not have an LLM backed queryable version of the docs?

[tansaku]

I want a queryable version of the langchain docs that I can ask things like:

• what default models are being used and how can I override them?
• how can I get query_with_sources to give me finer granularity into the segment of text the answer comes from?

and just other random stuff. I think I'm already spoiled by GPT4 being able to ask about systems without having to plough through docs. :-)

Given all the ask questions of text, is someone not already running a langchain bot that can answer questions about langchain?

Or is that just trivial two or three liner in a notebook? :-)

[tansaku]

from langchain.document_loaders import WebBaseLoader

loader = WebBaseLoader("https://python.langchain.com/en/latest/")
langchain_index = VectorstoreIndexCreator().from_loaders([loader])

query = "what default models are being used and how can I override them?"
langchain_index.query_with_sources(query)

Using embedded DuckDB without persistence: data will be transient
{'question': 'what default models are being used and how can I override them?',
'answer': ' LangChain supports various models, prompts, memory, indexes, and chains. It also provides custom agents, toolkits, agent executors, and use cases.\n',
'sources': 'https://python.langchain.com/en/latest/'}

[tansaku]

query = "how can I get query_with_sources to give me finer granularity into the segment of text the answer comes from?"
langchain_index.query_with_sources(query)

{'question': 'how can I get query_with_sources to give me finer granularity into the segment of text the answer comes from?',
'answer': ' You can use the Question Answering with Sources chain to get finer granularity into the segment of text the answer comes from.\n',
'sources': 'https://python.langchain.com/en/latest/'}

[tansaku]

so it seems like what I would want to do would involve combining the approach described here https://python.langchain.com/en/latest/modules/chains/index_examples/vector_db_qa_with_sources.html with the web document loader and some additional changes so that rather than sources being the the indices of the chunks, but the actual text itself ...

[tansaku]

so using a book text I have I can get answers like this:

{'answer': ' To stop bearing grudges, one can try to reach out to reconcile, focus on other things, and try to move on.\n',
 'sources': '427-pl, 424-pl'}

and so these number 427 and 424 appear to be indexing into the text I've chunked up

with open("book/contents.txt") as f:
    book = f.read()
text_splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=0)
texts = text_splitter.split_text(book)

but are they correctly 0 indexed? Given that I've got the texts list available I can look them up like texts[424] and texts[427] but could that be something that was just built into the chain itself? Aha, looks like there is a flag for RetrievalQAWithSourcesChain so

chain = RetrievalQAWithSourcesChain.from_chain_type(OpenAI(temperature=0), chain_type="stuff", retriever=docsearch.as_retriever(), return_source_documents=True)

which gives source text as well as ids, but strangely I get 'sources': '427-pl, 424-pl', but then four source documents which include 424, 427, 429 and 432 ... hmmm

[tansaku]

hmm, okay so there's https://github.com/hwchase17/chat-langchain but seems to be broken in various ways ...