March 2024 Plans

[hwchase17]

Hi all! Trying something new here. I wanted to share some of our priorities and roadmap for the upcoming month. We would love feedback on these issues - are we tackling the right things? Are there things that are missing? Are there places you would want to help?

Note that beside these main themes, we will also continue to focus on core things like integrations, chain examples, templates, bug fixes, LangSmith integration etc

Theme 1: Structured Outputs

• Continue to document usage of with_structured_output on LLMs
• Continue to add with_structured_output to more LLMs that support function calling
• Add benchmarks for extraction
• Improve extraction use case page
• Add extraction use case accelerant (standalone github repo)
• Add utilities for constructing few shot example messages
• Support streaming of structured outputs in more places
• Support retries with context

Theme 2: Retrieval

• Improve docs for query analysis
• Add query analysis use case accelerant (standalone github repo)
• Add abstraction for "index"

Theme 3: Agents

• Continue to add more examples of using LangGraph
• Specifically, add a few examples of complex chatbots

Theme 4: Documentation

• Add feedback buttons for each page
• Fix all broken links
• Based on feedback left on pages and other feedback, think about how to improve architecture of docs

Theme 5: Streaming improvements

• add async streaming implementations for more chat models
• make it easier to implement streaming when creating a custom chat model
• add more info to the output of astream_events()

[KBB99]

I'd love to see support for Claude 3 and see examples of using it with Langgraph. Link to feature discussion here.

[Nov1c444]

If LangChain Team could provide the following information, I will be very grateful:

1. Local LLM examples of LangGraph
2. The positioning of LangGraph in the overall architecture diagram of LangChain
3. The architecture diagram of LangGraph

[agustinsacco]

Really good example of agents using LangGraph and Ollama for generation and embeddings 👌

https://langchain-ai.github.io/langgraphjs/tutorials/rag/langgraph_adaptive_rag_local

[mingxuan-he]

Super excited for wider astream support! streaming events / logs in LangGraph is currently broken to a point where a large portion of example notebooks are not compatible with llm token streaming. Issues for reference: #78 #124 #136

[eyurtsev]

Thanks for pointing out, a fix will be out shortly

[Neykah]

Thanks for the great work!!!
Do you guys have an estimation for the release date of 0.2.0?
We still have a lot of deprecation warning to take care of in our codebase, and I'm wondering how urgent it is to tackle them all.

[ashburnham]

Great to see with_structured_output here 😎

One very important thing I think that has gone a bi backwards is the use of Zod in langchainjs

In the old code it uses Zod to parse the output, rather than using the zod schema to generate json. This is VERY important as you can use z.preprocess to handle simple (very common) LLM output errors BEFORE parsing to avoid loads of avoidable errors. e.g. when it's supposed to return an array, but it returns a long single string that that is a comm separated list of the elements, or when numbers come back as strings like "123".

https://github.com/colinhacks/zod?tab=readme-ov-file#preprocess

You can see how it used to be handled here

https://github.com/langchain-ai/langchainjs/blob/main/langchain/src/chains/openai_functions/structured_output.ts

In the new code, as far as I can tell, it just uses the zod schema to generate a json schema so so loses this rich functionality.

https://github.com/langchain-ai/langchainjs/blob/main/langchain/src/output_parsers/structured.ts

We can't (and I wouldn't recommend anyone else to) use structured outputs in production without this as the types of errors I describe are so common and so avoidable. They are especially common in Claude which doesn't do function calling so natively, and I know a lot of people are going to be using Claud with the recent releases.

I've submitted PRs for to make this work in the original version, but don't have time this month to do so. When 0.2.0 is released and the old code is dropped we won't be able to upgrade unless this functionality is maintained.

As with_structured_output is a priority, please can this feedback be taken on for 0.2.0?

I should probably start an issue shouldn't I... but I just saw it was the top priority here..

[jacoblee93]

Thanks for the feedback @ashburnham - don't see any reason why we couldn't make this the default as it's a beta API, and should be the default for implementations in general.

Could you open an issue?

[jacoblee93]

I fixed this today for all the client libraries! Please update and try it out. Thank you for the report!

[Nazareka]

As I see in this discussion, it would be nice to add a disclaimer about Jupiter Notebook. Because it's not obvious for some new guys.

[sepiatone]

This is already present, under Get Started / QuickStart, added late last year via #15281. Any suggestions on how this could be made more prominent?

[xxshubhamxx]

Can't wait for more models to be added for with_structured_output. It is an incredibly useful functionality.

[Bongomannn]

Hello thanks for your work at all. How can i achieve here a decision between one or more tools with the new syntax:
llm.with_structured_output(takes_only_one_BaseClass_but_i_have_100_other_classes)