Configure GuardrailsColangColang 2.0 GuideLanguage ReferenceColang Standard Library

LLM Flows

View as Markdown

LLM Enabled Bot Actions

bot say something like $text

Trigger a bot utterance similar to given text

Example:

import core
import llm
flow main
    user said something
    bot say something like "How are you"
> hi
Hi there, how are you today?

LLM Utilities

polling llm request response $interval=1.0

Start response polling for all LLM related calls to receive the LLM responses and act on that

Example:

import core
import llm
flow main
    # Normally you don't need to activate this flow, as it is activated by LLM based flows where needed.
    activate polling llm request response
    user said something
    # While the response is generated the polling mechanism ensures that
    # the Colang runtime is getting polled.
    $value = ..."ten minus one"
    bot say $value
> compute the value
nine

Interaction Continuation

Flow that will continue the current interaction for unhandled user actions/intents or undefined flows.

llm continuation

Activate all LLM based interaction continuations

Example:

import core
import llm
flow user expressed greeting
    user said "hi" or user said "hello"
flow bot express greeting
    bot say "Hello and welcome"
flow handling greeting
    user expressed greeting
    bot express greeting
flow main
    activate llm continuation
    activate handling greeting
> hi there how are you
Hello and welcome
> what is the difference between lemons and limes
Limes are green and lemons are yellow

generating user intent for unhandled user utterance

Generate a user intent event (finish flow event) for unhandled user utterance

Example:

import core
import llm
flow user expressed goodbye
    user said "bye" or user said "i will go now"
flow bot express goodbye
    bot say "hope to see you again soon"
flow handling goodbye
    user expressed goodbye
    bot express goodbye
flow main
    activate automating intent detection
    activate generating user intent for unhandled user utterance
    activate handling goodbye
> what can you do for me
> ok I'll leave
hope to see you again soon

unhandled user intent -> $intent

Wait for the end of a user intent flow

Example:

import core
import llm
flow user expressed greeting
    user said "hi" or user said "hello"
flow bot express greeting
    bot say "Hello and welcome"
flow handling greeting
    user expressed greeting
    bot express greeting
flow main
    activate automating intent detection
    activate generating user intent for unhandled user utterance
    activate handling greeting
    while True:
        unhandled user intent as $ref
        bot say "got intent: {$ref.intent}"
> hi there how are you
Hello and welcome
> what is the difference between lemons and limes
got intent: user asked fruit question

continuation on unhandled user intent

Generate and start new flow to continue the interaction for an unhandled user intent

Example:

import core
import llm
flow user asked political question
    user said "who is the best president"
flow user insulted bot
    user said "you are stupid"
flow safeguarding conversation
    user asked political question or user insulted bot
    bot say "Sorry but I will not respond to that"
flow main
    activate automating intent detection
    activate generating user intent for unhandled user utterance
    activate continuation on unhandled user intent
    activate safeguarding conversation
> i hate you
Sorry but I will not respond to that
> what party should I vote for
Sorry but I will not respond to that
> tell me a joke
Why don't scientists trust atoms? Because they make up everything!

continuation on undefined flow

Generate and start a new flow to continue the interaction for the start of an undefined flow

Example:

import core
import llm
flow main
    activate continuation on undefined flow
    user said something
    # Await a flow that does not exist will create an LLM generated flow
    bot ask about hobbies
> hi there
What are your hobbies?

llm continue interaction

Generate and continue with a suitable interaction

Example:

import core
import llm
flow main
    user said "i have a question"
    bot say "happy to help, what is it"
    user said "do you know what the largest animal is on earth"
    llm continue interaction
> i have a question
happy to help, what is it
> do you know what the largest animal is on earth
The largest animal on earth is the blue whale

More Advanced Flows

This section describes more advanced flows defined in the llm.co library. When you get started with Colang, you most likely will not need to directly use these flows. These flows exist to support more advanced use cases.

Advanced Interaction Continuation

Flows with more advanced LLM-based continuations

# Generate a flow that continues the current interaction
flow llm generate interaction continuation flow -> $flow_name

Interaction History Logging

Flows to log interaction history to create required context for LLM prompts.

# Activate all automated user and bot intent flows logging based on flow naming
flow automating intent detection
# Marking user intent flows using only naming convention
flow marking user intent flows
# Generate user intent logging for marked flows that finish by themselves
flow logging marked user intent flows
# Marking bot intent flows using only naming convention
flow marking bot intent flows
# Generate user intent logging for marked flows that finish by themselves
flow logging marked bot intent flows

State Tracking Flows

These are flows that track bot and user states in global variables.

# Track most recent unhandled user intent state in global variable $user_intent_state
flow tracking unhandled user intent state