New Feature - Output Parsers

[vinodkiran]

@HenryHengZJ Need your (early) feedback on this new Feature.

Lanchain Output Parsers

Very initial (draft) implementation for handing output parsers in Flowise.

Do not spend time reviewing the code, it needs a lot of cleanup / refactoring.

[vinodkiran]

[HenryHengZJ]

awesome! will use this as the base for further improvement.

1.) In this PR we will be returning the reponse as a JSON instead of string - { text: 'some output' }. If we have output parser, would prefer to have smtg like { json: {key: val} }. And from the UI, we can parse it and return as a nice JSON UI;

2.) For Zod Schema, would love to have a UI that allow easier schema input. Similar to custom tool:

3.) From the PR, we have if else check to cater for PromptTemplate and ChatPromptTemplate, what about other prompt like FewShot, can we use output parser for that, if not, how do we disable connection

Just few thoughts, don't have to implement all of these yet

[vinodkiran]

3. Fewshot and other templates: Yes, I am working on this. was eager to get some early directional feedback, so I created PR. I will push a new commit with changes to handle the other template types.

1 & 2): Could be a new PR, as they will add a lot of value. After the 1.0 release of this component node.

[vinodkiran]

[HenryHengZJ]

looks great! are you going to add more output parser? let me know when its ready to review

[HenryHengZJ]

@vinodkiran would love to have the 1.) implemented for the next PR, so we can properly return the JSON when calling API, and this can lead to us starting to implement extraction/tagging chain - https://js.langchain.com/docs/modules/chains/popular/structured_output. What do you think?

[vinodkiran]

@vinodkiran would love to have the 1.) implemented for the next PR, so we can properly return the JSON when calling API, and this can lead to us starting to implement extraction/tagging chain - https://js.langchain.com/docs/modules/chains/popular/structured_output. What do you think?

@HenryHengZJ Agreed. Makes perfect sense. I will update this PR to return json...when will the base PR be merged?

[HenryHengZJ]

will be testing it today

[vinodkiran]

will be testing it today

Sporadically, I see double responses to the front end. Also, sometimes the response is shown before the output parser (parse method) is invoked (in a debug mode). Kindly validate.

[HenryHengZJ]

you mean this?

seems to happen when running yarn dev, yarn start works fine though

[HenryHengZJ]

@vinodkiran can we use this as base branch for Output Parser, and create branch for further PRs and merge into this. Once everything is in, then we merge to main and release to public.

[HenryHengZJ]

@vinodkiran can you allow edits from maintainer on your fork? here's guide

[vinodkiran]

@vinodkiran can you allow edits from maintainer on your fork? here's guide

Yes, maintainer edits are enabled for this PR.

[vinodkiran]

@HenryHengZJ we could probably use the JsonStructuredOutputParser and not hardcode ```json in the downstream classes.

[HenryHengZJ]

Update: Changed StructuredOutputParser to use Zod Schema:

[HenryHengZJ]

@vinodkiran let me know if you have any question before merging

[vinodkiran]

@vinodkiran let me know if you have any question before merging

@HenryHengZJ 1) do we continue to use the word experimental in the category? 2) currently the input parameter is defined for LLMchain only and not for others. Do we target the rest as a separate PR?

[HenryHengZJ]

@vinodkiran let me know if you have any question before merging

@HenryHengZJ 1) do we continue to use the word experimental in the category? 2) currently the input parameter is defined for LLMchain only and not for others. Do we target the rest as a separate PR?

1. Yeah we can remove the word experimental
2. I think for starter we can have output parser only for Llmchain. And we can add them to other chains if there is further demand or requests

[HenryHengZJ]

Merged in #1009, resolving the conflict now