switch-case over free-form sentences.
Project description
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
'''
- Set up a
LangStore
data container.
data = yaml.safe_load(triggers)
S = LangStore(data)
- 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
returnsNone
if the nearest topic distance is greater than the threshold.
Debug and More!
- Use
debug=True
anddebug_k=5
to- display distance of
user_msg
to different topics. - Get match scores for each message group to debug what went wrong.
- display distance of
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 graph-based Conversation Flow
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 modules
- more control over conversation flow, without losing ability to chitchat
- build conversations, not (worry about) LLM chains!
For example, see a sales lead filtering and appointment-booking bot under examples.
- Specify the assistant and user messages as well as the overall conversation flow in a yaml file easily
- Automate the conversational flow with few lines of code
- Add messages and modify the flow effortlessly by modifying the yaml
- Model picks up the changes instantly, no need to fine-tune models for the changes
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.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for langchoice-0.0.3-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | cb17ee1d78b37fbdd3aa47fc74c53477aeb61431a90476a3184f5bbe65067469 |
|
MD5 | 4608c1f4a0ed1ff7d05feb7ee4fb4423 |
|
BLAKE2b-256 | 2e87f024094f1c0ce6fe6b9a8e61592487b27de59133037bcf3a862280e32a8c |