refactor(llm): remove LangChain imports from core modules

nemoguardrails/actions/action_dispatcher.py

@@ -23,8 +23,6 @@
 from pathlib import Path
 from typing import Any, Callable, Dict, List, Optional, Tuple, Type, Union, cast
 
-from langchain_core.runnables import Runnable
-
 from nemoguardrails import utils
 from nemoguardrails.exceptions import LLMCallException
 
@@ -218,11 +216,10 @@ async def execute_action(
                         else:
                             log.warning(f"Synchronous action `{action_name}` has been called.")
 
-                    elif isinstance(fn, Runnable):
-                        # If it's a Runnable, we invoke it as well
-                        runnable = fn
-
-                        result = await runnable.ainvoke(input=params)
+                    elif hasattr(fn, "ainvoke") and callable(fn.ainvoke):  # type: ignore[union-attr]
+                        # Duck-type check for LangChain Runnables (or any object
+                        # with ainvoke) to avoid importing langchain in core.
+                        result = await fn.ainvoke(input=params)  # type: ignore[union-attr]
                     else:
                         # TODO: there should be a common base class here
                         fn_run_func = getattr(fn, "run", None)

nemoguardrails/eval/cli.py

@@ -19,8 +19,6 @@
 from typing import List
 
 import typer
-from langchain_community.cache import SQLiteCache
-from langchain_core.globals import set_llm_cache
 
 from nemoguardrails.eval.check import LLMJudgeComplianceChecker
 from nemoguardrails.eval.eval import run_eval
@@ -172,8 +170,14 @@ def check_compliance(
     if disable_llm_cache:
         console.print("[orange]Caching is disabled.[/]")
     else:
-        console.print("[green]Caching is enabled.[/]")
-        set_llm_cache(SQLiteCache(database_path=".langchain.db"))
+        try:
+            from langchain_community.cache import SQLiteCache
+            from langchain_core.globals import set_llm_cache
+
+            set_llm_cache(SQLiteCache(database_path=".langchain.db"))
+            console.print("[green]Caching is enabled.[/]")
+        except ImportError:
+            console.print("[yellow]langchain not installed, LLM caching unavailable.[/]")
 
     console.print(f"Using eval configuration from {eval_config_path}.")
     console.print(f"Using output paths: {output_path}.")

nemoguardrails/evaluate/evaluate_factcheck.py

@@ -20,14 +20,14 @@
 
 import tqdm
 import typer
-from langchain_core.prompts import PromptTemplate
 
 from nemoguardrails import LLMRails
 from nemoguardrails.actions.llm.utils import llm_call
 from nemoguardrails.evaluate.utils import load_dataset
 from nemoguardrails.llm.prompts import Task
 from nemoguardrails.llm.taskmanager import LLMTaskManager
 from nemoguardrails.rails.llm.config import RailsConfig
+from nemoguardrails.utils import get_or_create_event_loop
 
 
 class FactCheckEvaluation:
@@ -89,27 +89,19 @@ def create_negative_samples(self, dataset):
         that it will not be grounded in the evidence passage. change details in the answer to make the answer
         wrong but yet believable.\nevidence: {evidence}\nanswer: {answer}\nincorrect answer:"""
 
-        create_negatives_prompt = PromptTemplate(
-            template=create_negatives_template,
-            input_variables=["evidence", "answer"],
-        )
-
-        # Bind config parameters to the LLM for generating negative samples
-        llm_with_config = self.llm.bind(temperature=0.8, max_tokens=300)
+        loop = get_or_create_event_loop()
 
         print("Creating negative samples...")
         for data in tqdm.tqdm(dataset):
             assert "evidence" in data and "question" in data and "answer" in data
             evidence = data["evidence"]
             answer = data["answer"]
 
-            # Format the prompt and invoke the LLM directly
-            formatted_prompt = create_negatives_prompt.format(evidence=evidence, answer=answer)
-            negative_answer = llm_with_config.invoke(formatted_prompt)
-            if isinstance(negative_answer, str):
-                data["incorrect_answer"] = negative_answer.strip()
-            else:
-                data["incorrect_answer"] = negative_answer.content.strip()
+            formatted_prompt = create_negatives_template.format(evidence=evidence, answer=answer)
+            response = loop.run_until_complete(
+                self.llm.generate_async(formatted_prompt, temperature=0.8, max_tokens=300)
+            )
+            data["incorrect_answer"] = response.content.strip()
 
         return dataset
 

nemoguardrails/library/hallucination/actions.py

@@ -17,8 +17,6 @@
 import logging
 from typing import Optional
 
-from langchain_core.prompts import PromptTemplate
-
 from nemoguardrails import RailsConfig
 from nemoguardrails.actions import action
 from nemoguardrails.actions.llm.utils import (
@@ -56,8 +54,7 @@ async def self_check_hallucination(
     if bot_response and last_bot_prompt_string:
         num_responses = HALLUCINATION_NUM_EXTRA_RESPONSES
 
-        last_bot_prompt = PromptTemplate(template="{text}", input_variables=["text"])
-        formatted_prompt = last_bot_prompt.format(text=last_bot_prompt_string)
+        formatted_prompt = last_bot_prompt_string
 
         async def _generate_extra_response(index: int) -> Optional[str]:
             llm_call_info_var.set(LLMCallInfo(task=Task.SELF_CHECK_HALLUCINATION.value))