engine.py

# SPDX-FileCopyrightText: Copyright (c) 2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.


import logging

from langchain.agents import AgentType
from langchain.agents import Tool
from langchain.agents import initialize_agent
from langchain.chains.retrieval_qa.base import RetrievalQA
from langchain.vectorstores.faiss import FAISS
from langchain_core.embeddings import Embeddings

from morpheus_llm.llm import LLMContext
from morpheus_llm.llm import LLMEngine
from morpheus_llm.llm.nodes.extracter_node import ManualExtracterNode
from morpheus_llm.llm.services.llm_service import LLMService
from morpheus_llm.llm.services.utils.langchain_llm_client_wrapper import LangchainLLMClientWrapper
from morpheus_llm.llm.task_handlers.simple_task_handler import SimpleTaskHandler

from ..data_models.config import RunConfig
from ..data_models.info import AgentMorpheusInfo
from ..nodes.cve_checklist_node import CVEChecklistNode
from ..nodes.cve_justification_node import CVEJustifyNode
from ..nodes.cve_langchain_agent_node import CVELangChainAgentNode
from ..nodes.cve_summary_node import CVESummaryNode
from ..utils.code_searcher import LangchainCodeSearcher
from ..utils.document_embedding import DocumentEmbedding
from ..utils.prompting import agent_examples_for_prompt
from ..utils.serp_api_wrapper import MorpheusSerpAPIWrapper

logger = logging.getLogger(__name__)


def _build_dynamic_agent_fn(run_config: RunConfig, embeddings: Embeddings):

    chat_service = LLMService.create(run_config.engine.agent.model.service.type,
                                     **run_config.engine.agent.model.service.model_dump(exclude={"type"},
                                                                                        by_alias=True))
    chat_client = chat_service.get_client(**run_config.engine.agent.model.model_dump(exclude={"service", "type"},
                                                                                     by_alias=True))
    langchain_llm = LangchainLLMClientWrapper(client=chat_client)

    # Initialize a SerpAPIWrapper object to perform internet searches.
    search = MorpheusSerpAPIWrapper(max_retries=run_config.general.max_retries)

    # Append new Tools to the tools list, which allows for internet searches and software version comparisons.
    # The first tool can be especially useful for answering questions about external libraries while the second
    # allows for more consistent and accurate comparisons of software versions.

    def inner_create_agent_fn(context: LLMContext):

        tools: list[Tool] = [
            Tool(
                name="Internet Search",
                func=search.run,  # Synchronous function for running searches.
                coroutine=search.arun,  # Asynchronous coroutine for running searches.
                description="useful for when you need to answer questions about external libraries",
            ),
        ]

        vdb_map: AgentMorpheusInfo.VdbPaths = context.message().get_metadata("info.vdb")  # type: ignore

        def run_retrieval_qa_tool(retrieval_qa_tool: RetrievalQA, query: str) -> str | dict:
            """
            Runs a given retrieval QA tool on the provided query. Returns a dict of the result string and source
            documents if the `return_source_documents` config is true, otherwise it returns just the result string if
            `return_source_documents` is false.
            """
            output_dict = retrieval_qa_tool(query)

            # If returning source documents, include the result and source_documents keys in the output
            if run_config.engine.agent.return_source_documents:
                return {k: v for k, v in output_dict.items() if k in ["result", "source_documents"]}

            # If not returning source documents, return only the result as a string
            else:
                return output_dict["result"]

        if (vdb_map.code_vdb_path is not None):
            # load code vector DB
            code_vector_db = FAISS.load_local(vdb_map.code_vdb_path, embeddings, allow_dangerous_deserialization=True)
            code_qa_tool = RetrievalQA.from_chain_type(
                llm=langchain_llm,
                chain_type="stuff",
                retriever=code_vector_db.as_retriever(),
                return_source_documents=run_config.engine.agent.return_source_documents)
            tools.append(
                Tool(name="Docker Container Code QA System",
                     func=lambda query: run_retrieval_qa_tool(code_qa_tool, query),
                     description=("useful for when you need to check if an application or any dependency within "
                                  "the Docker container uses a function or a component of a library.")))
        elif run_config.general.code_search_tool:

            logger.info("Preparing source code documents for the code search tool.")

            # Use existing document loader and chunker from DocumentEmbedding class, without embedding.
            embedder = DocumentEmbedding(embedding=None,
                                         vdb_directory=run_config.general.base_vdb_dir,
                                         git_directory=run_config.general.base_git_dir)

            documents = []
            sources = context.message().get_metadata("input").image.source_info
            for source_info in sources:
                if source_info.type == 'code':
                    documents.extend(embedder.collect_documents(source_info))

            if len(documents) > 0:
                documents_index = embedder._chunk_documents(documents)
                lexical_code_searcher = LangchainCodeSearcher(documents_index, rank_documents=True, k=5)

                tools.append(
                    Tool(name="Docker Container Code Search",
                         func=lexical_code_searcher.search,
                         description=("useful for when you need to search the Docker container's code for a given "
                                      "function or component of a library. This requires exact function name or library"
                                      "without no additional information")))
            else:
                logger.warning("No code documents found for the code search tool.")

        if (vdb_map.doc_vdb_path is not None):
            guide_vector_db = FAISS.load_local(vdb_map.doc_vdb_path, embeddings, allow_dangerous_deserialization=True)
            guide_qa_tool = RetrievalQA.from_chain_type(
                llm=langchain_llm,
                chain_type="stuff",
                retriever=guide_vector_db.as_retriever(),
                return_source_documents=run_config.engine.agent.return_source_documents)
            tools.append(
                Tool(name="Docker Container Developer Guide QA System",
                     func=lambda query: run_retrieval_qa_tool(guide_qa_tool, query),
                     description=(
                         "Useful for when you need to ask questions about the purpose and functionality of the Docker "
                         "container.")))

        # Define a system prompt that sets the context for the language model's task. This prompt positions the assistant
        # as a powerful entity capable of investigating CVE impacts on Docker containers.
        sys_prompt = (
            "You are a very powerful assistant who helps investigate the impact of reported Common Vulnerabilities and "
            "Exposures (CVE) on Docker containers. Information about the Docker container under investigation is stored in "
            "vector databases available to you via tools.")

        # Initialize an agent with the tools and settings defined above.
        # This agent is designed to handle zero-shot reaction descriptions and parse errors.
        agent = initialize_agent(
            tools,
            langchain_llm,
            agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION,
            verbose=run_config.engine.agent.verbose,
            handle_parsing_errors="Check your output and make sure it conforms, use the Action/Action Input syntax",
            max_iterations=10,
            early_stopping_method="generate",
            return_intermediate_steps=run_config.engine.agent.return_intermediate_steps)

        # Modify the language model chain's prompt template to adjust how the model should process inputs and structure
        # responses.
        prompt_template = agent.agent.llm_chain.prompt.template.replace(
            "Answer the following questions as best you can.",
            ("If the input is not a question, formulate it into a question first. Include intermediate thought in the "
             "final answer."),
            1).replace(
                "Use the following format:",
                "Use the following format (start each response with one of the following prefixes: "
                "[Question, Thought, Action, Action Input, Final Answer]):",
                1)
        if run_config.engine.agent.prompt_examples:
            prompt_template = prompt_template.replace("Begin!\n\n", agent_examples_for_prompt + "Begin!\n\n")
        agent.agent.llm_chain.prompt.template = f'{sys_prompt} {prompt_template}'

        return agent

    return inner_create_agent_fn


def build_engine(*, run_config: RunConfig, embeddings: Embeddings):

    summary_service = LLMService.create(run_config.engine.summary_model.service.type,
                                        **run_config.engine.summary_model.service.model_dump(exclude={"type"}))
    justification_service = LLMService.create(
        run_config.engine.justification_model.service.type,
        **run_config.engine.justification_model.service.model_dump(exclude={"type"}))

    engine = LLMEngine()

    checklist_node = CVEChecklistNode(checklist_model_config=run_config.engine.checklist_model,
                                      enable_llm_list_parsing=run_config.general.enable_llm_list_parsing)

    engine.add_node("extract_prompt", node=ManualExtracterNode(input_names=checklist_node.get_input_names()))

    engine.add_node("checklist", inputs=[("/extract_prompt/*", "*")], node=checklist_node)

    engine.add_node("agent",
                    inputs=[("/checklist", "input")],
                    node=CVELangChainAgentNode(
                        create_agent_executor_fn=_build_dynamic_agent_fn(run_config, embeddings),
                        replace_exceptions=True,
                        replace_exceptions_value="I do not have a definitive answer for this checklist item."))

    engine.add_node('summary',
                    inputs=[("/checklist", "checklist_inputs"), ("/agent/outputs", "checklist_outputs"),
                            "/agent/intermediate_steps"],
                    node=CVESummaryNode(llm_client=summary_service.get_client(
                        **run_config.engine.summary_model.model_dump(exclude={"service", "type"}))))

    engine.add_node('justification',
                    inputs=[("/summary/summary", "summaries")],
                    node=CVEJustifyNode(llm_client=justification_service.get_client(
                        **run_config.engine.justification_model.model_dump(exclude={"service", "type"}))))

    handler_inputs = [
        "/summary/checklist",
        "/summary/summary",
        f"/justification/{CVEJustifyNode.JUSTIFICATION_LABEL_COL_NAME}",
        f"/justification/{CVEJustifyNode.JUSTIFICATION_REASON_COL_NAME}",
        f"/justification/{CVEJustifyNode.AFFECTED_STATUS_COL_NAME}",
    ]
    handler_outputs = [
        "checklist",
        "summary",
        CVEJustifyNode.JUSTIFICATION_LABEL_COL_NAME,
        CVEJustifyNode.JUSTIFICATION_REASON_COL_NAME,
        CVEJustifyNode.AFFECTED_STATUS_COL_NAME
    ]

    # Add our task handler
    engine.add_task_handler(inputs=handler_inputs, handler=SimpleTaskHandler(output_columns=handler_outputs))

    return engine