Skip to content

"if-then-else" over topics made up of free-form sentences. Build conversations, not LLM chains!

License

Notifications You must be signed in to change notification settings

ekshaks/langchoice

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

14 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

License PyPi

PyPI: https://pypi.org/project/langchoice

🤖 About LangChoice

switch-case, but for free-form sentences. A one-liner for if-then-else over similar sentences.

The LangChoice library allows you to condition your structured programs over natural language sentence predicates or triggers. Makes it easy to define conditional flows over user inputs without implementing the sentence match operator over and over again.

📦 Installation

pip install langchoice

🧪 Getting Started

Suppose we want to detect if an incoming user message triggered belongs to one of the following (greeting, politics) topics. Do this in a few lines using langchoice.

First define the text messages that define each topic category (a message group).

triggers =\
'''
greeting:
    - hello
    - hi
    - what's up?

politics:
    - what are your political beliefs?
    - thoughts on the president?
    - left wing
    - right wing
'''
  1. Set up a LangStore data container.
data = yaml.safe_load(triggers)
S = LangStore(data)
  1. On receiving a user message user_msg, simply match with topics!
match S.match(user_msg, threshold=1.2, debug=True): 
    case 'greeting', _ : #user_msg matches the greeting message cluster
        say_hello()
    case 'politics', _ : #user_msg matches the politics message cluster
        change_topic()
    case x :
        print(x)
        print(f'No defined triggers detected. Ask an LLM for response.')

Add or remove sentences from each topic or introduce new topics. Works on-the-fly!

Supports multiple matching modes:

  • S.match returns the topic of nearest message.
  • Use S.match_centroid to instead find the nearest topic centroid .
  • (Optional) Provide thresholds for the no-topic-matches case:
    • S.match returns None if the nearest topic distance is greater than the threshold.

Debug and More!

  • Use debug=True and debug_k=5 to
    • display distance of user_msg to different topics.
    • Get match scores for each message group to debug what went wrong.

Coming Soon!

  • In built assertions, which fail if the execution fails to match the expected topic / group.
  • Compute the thresholds automatically for pre-specified message groups against a query evaluation set
  • Fine-tune embeddings to separate message clusters better.

Implement a Conversation Flow with Message-based Routing

The langchoice package enables you to make controlled chatbot flows as well as build guardrails very quickly.

The key motivation is to allow users to have maximal control when designing the bot:

  • add/update messages on demand
  • route user messages to owner states or modules
  • more control over conversation flow, without losing ability to chitchat
  • build conversations, not (worry about) LLM chains!

Check out the sales lead filtering and appointment-booking bots under examples.

Roadmap

  • match variants
    • hybrid match - match regex separately
    • local S.match (based on state tag)
    • match type='item' (default) | type='group' (centroid) | ...
  • allow switching under-the-hood sentence encoders
  • fine-tune based on in-topic and irrelevant messages

Author

Nishant Sinha, Founder, Consulting Scientist, OffNote Labs.

For insights on Generative AI evolution and applications, follow on Linkedin, and read our articles on Substack.