feat: support langchain v1

[Pouyanpi]

PR Description

Summary

Adds full LangChain 1.x support while maintaining backward compatibility with 0.x versions. This PR migrates all imports to use canonical langchain-core paths, removes deprecated features (Chain support, LLMParams, SummarizeDocument), and provides documentation updates.

BREAKING CHANGES:

• Chain support removed from action dispatcher (use Runnable instead)
• SummarizeDocument built-in action removed (implement custom version)
• langchain-nvidia-ai-endpoints removed from optional dependencies

Key Changes

Dependency Updates

• Extended langchain constraint from <0.4.0 to <2.0.0
• Extended langchain-core constraint from <0.4.0 to <2.0.0
• Extended langchain-community constraint from <0.4.0 to <2.0.0

Import Standardization (28 files)

Migrated all imports to canonical paths:

# before (deprecated)
from langchain.chat_models.base import BaseChatModel
from langchain_core.language_models.llms import BaseLLM

# after (correct)
from langchain_core.language_models import BaseChatModel, BaseLLM

Removed Deprecated Features

• Chain support : Removed from action_dispatcher.py (use Runnable instead)
• LLMParams : Removed context manager (use native LLM params)
• SummarizeDocument : Removed built-in action (implement custom)
• langchain-nvidia-ai-endpoints : Removed from optional deps (install manually untill it supports v1)

Documentation Updates

• Completely rewrote custom LLM provider guide
• Added BaseChatModel documentation (was missing!)
• Added migration guide for Chain → Runnable
• [ ] update CHANGELOG.md with breaking changes

Backward Compatibility

Examples include fallback pattern for LangChain 1.x users:

try:
    from langchain.chains import RetrievalQA
except ImportError:
    from langchain_classic.chains import RetrievalQA

[github-actions[bot]]

Documentation preview

https://nvidia-nemo.github.io/Guardrails/review/pr-1472

[codecov-commenter]

Codecov Report

:x: Patch coverage is 64.00000% with 18 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
nemoguardrails/evaluate/evaluate_factcheck.py	10.00%	9 Missing :warning:
nemoguardrails/library/hallucination/actions.py	33.33%	4 Missing :warning:
...moguardrails/llm/providers/huggingface/pipeline.py	0.00%	3 Missing :warning:
nemoguardrails/llm/helpers.py	0.00%	2 Missing :warning:

:loudspeaker: Thoughts on this report? Let us know!

[Pouyanpi]

@greptileai

[Pouyanpi]

Summary of ALL LangChain Imports in Codebase:

1. langchain_core imports (All CORRECT) from langchain_core.language_models import BaseLLM, BaseChatModel, BaseLanguageModel, LLM from langchain_core.callbacks.manager import ... from langchain_core.messages import ... from langchain_core.outputs import ... from langchain_core.prompts import ... from langchain_core.runnables import ... from langchain_core.tools import ... All are canonical paths from langchain-core

2. langchain_community imports (All CORRECT) from langchain_community import llms from langchain_community.cache import SQLiteCache from langchain_community.chat_models import _module_lookup from langchain_community.llms import HuggingFacePipeline All are correct community package imports

3. langchain (Main Package) - (✓ CORRECT) from langchain.chat_models import init_chat_model Confirmed: init_chat_model STAYS in langchain.chat_models in v1.0

• It's part of the "streamlined core namespace"
• Provides essential model initialization

4. langchain_nvidia_ai_endpoints (✓ CORRECT) from langchain_nvidia_ai_endpoints import ChatNVIDIA Separate integration package, correct

all import paths are correct and V1.0 compatible!

No additional changes needed. The codebase is:

• Using canonical langchain-core paths
• Using correct langchain_community paths
• Using correct init_chat_model from main package (stays in v1.0)
• No deprecated proxy imports remaining
• Fully compatible with langchain 0.2.x, 0.3.x, and 1.x

The only import from the main langchain package (init_chat_model) is intentionally kept there in v1.0 as part of the core functionality. Everything else correctly imports from langchain-core or langchain-community.

[greptile-apps[bot]]

Greptile Overview

Greptile Summary

This PR successfully migrates NeMo Guardrails to support LangChain v1.x while maintaining backward compatibility with v0.x. The migration standardizes imports to canonical langchain_core paths, removes deprecated Chain-based patterns in favor of Runnables and direct LLM invocation, and updates 28 files across the codebase.

Key Changes:

• Removed Chain support from action dispatcher (replaced with Runnable pattern)
• Migrated from LLMChain to llm.bind() for runtime parameter configuration
• Standardized imports: from langchain_core.language_models import BaseLLM, BaseChatModel
• Updated model initialization to use LangChain v1's init_chat_model() canonical API
• Added langchain-classic fallback pattern in examples for v1.x users
• Removed langchain-nvidia-ai-endpoints from optional dependencies (users must install manually until v1 support)

Issues Found:

• Version constraints allow untested LangChain 2.x (<2.0.0 should be <1.0.0)
• Test file custom_chat_model.py incorrectly implements _call()/_acall() instead of required _generate() methods for BaseChatModel

Confidence Score: 3/5

• This PR requires attention before merging due to overly permissive version constraints and a broken test implementation
• Score reflects well-executed migration patterns (Chain to Runnable, canonical imports) offset by two critical issues: (1) version constraints allowing untested LangChain 2.x releases that will introduce breaking changes, and (2) a test file implementing the wrong interface for BaseChatModel that will fail at runtime
• Pay close attention to pyproject.toml (version constraints) and tests/test_configs/with_custom_chat_model/custom_chat_model.py (incorrect interface implementation)

Important Files Changed

File Analysis

Filename	Score	Overview
pyproject.toml	3/5	Updated LangChain dependencies from <0.4.0 to <2.0.0, which is overly permissive and may introduce untested breaking changes from v2.x
nemoguardrails/actions/action_dispatcher.py	5/5	Cleanly removed deprecated Chain support, retained Runnable support for LangChain v1 compatibility
nemoguardrails/llm/models/langchain_initializer.py	5/5	Well-structured model initialization with proper fallbacks and error handling using canonical LangChain v1 imports
nemoguardrails/library/hallucination/actions.py	4/5	Migrated from LLMChain to direct llm.bind() pattern, correctly uses model_fields for Pydantic v2 and passes callbacks.handlers
tests/test_configs/with_custom_chat_model/custom_chat_model.py	2/5	Updated imports to canonical paths but implements _call/_acall methods instead of required _generate/_agenerate for BaseChatModel

Sequence Diagram

sequenceDiagram
    participant User
    participant LLMRails
    participant ActionDispatcher
    participant LangChainInit
    participant LLM as LLM/ChatModel
    participant Action as Custom Actions

    Note over User,Action: LangChain v1 Migration Flow

    User->>LLMRails: Initialize with config
    LLMRails->>LangChainInit: init_langchain_model(provider, model, mode)
    
    alt Chat Model Initialization
        LangChainInit->>LangChainInit: Try init_chat_model() (v1 canonical)
        LangChainInit->>LLM: Returns BaseChatModel instance
    else Text Completion Model
        LangChainInit->>LangChainInit: Try text completion providers
        LangChainInit->>LLM: Returns BaseLLM instance
    end

    LangChainInit-->>LLMRails: Initialized LLM instance
    
    Note over LLMRails,ActionDispatcher: Removed: Chain support
    
    User->>LLMRails: Execute action
    LLMRails->>ActionDispatcher: execute_action(name, params)
    
    alt Runnable Action (v1 pattern)
        ActionDispatcher->>Action: Check isinstance(fn, Runnable)
        ActionDispatcher->>Action: await runnable.ainvoke(params)
        Action-->>ActionDispatcher: Result
    else Function/Method Action
        ActionDispatcher->>Action: fn(**params)
        Action->>LLM: llm.bind(temperature=X, n=Y)
        Note over Action,LLM: Use bind() instead of Chain.with_config()
        Action->>LLM: await llm.agenerate(prompts, callbacks=handlers)
        LLM-->>Action: LLMResult with generations
        Action-->>ActionDispatcher: Result
    end
    
    ActionDispatcher-->>LLMRails: Result
    LLMRails-->>User: Response