|
| 1 | +from flask import Flask |
| 2 | +from flask import request, jsonify |
| 3 | +import threading |
| 4 | +from langchain_openai import OpenAIEmbeddings |
| 5 | +from langchain.chains import RetrievalQAWithSourcesChain, LLMChain |
| 6 | +import os |
| 7 | +from langchain.memory import ConversationBufferWindowMemory |
| 8 | +from langchain_community.vectorstores import Chroma |
| 9 | +from langchain_openai import OpenAI |
| 10 | +from langchain_community.document_loaders import DirectoryLoader |
| 11 | +from langchain.memory import ConversationBufferWindowMemory |
| 12 | +from langchain.text_splitter import CharacterTextSplitter |
| 13 | +from langchain_core.prompts import PromptTemplate |
| 14 | +from langchain import PromptTemplate |
| 15 | +from langchain_community.document_loaders import UnstructuredMarkdownLoader |
| 16 | +import logging |
| 17 | + |
| 18 | +app = Flask(__name__) |
| 19 | + |
| 20 | +retriever = None |
| 21 | +persist_directory = os.environ.get("PERSIST_DIRECTORY") |
| 22 | +vulnerable_app_qa = None |
| 23 | +target_source_chunks = int(os.environ.get("TARGET_SOURCE_CHUNKS", 4)) |
| 24 | +loaded_model_lock = threading.Lock() |
| 25 | +loaded_model = False |
| 26 | +logger = logging.getLogger(__name__) |
| 27 | +logger.setLevel(logging.DEBUG) |
| 28 | + |
| 29 | + |
| 30 | +def document_loader(): |
| 31 | + try: |
| 32 | + load_dir = "retrieval" |
| 33 | + logger.debug("Loading documents from %s", load_dir) |
| 34 | + loader = DirectoryLoader( |
| 35 | + load_dir, glob="**/*.md", loader_cls=UnstructuredMarkdownLoader |
| 36 | + ) |
| 37 | + documents = loader.load() |
| 38 | + logger.debug("Loaded %s documents", len(documents)) |
| 39 | + text_splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=100) |
| 40 | + texts = text_splitter.split_documents(documents) |
| 41 | + embeddings = get_embeddings() |
| 42 | + os.system("rm -rf ./db") |
| 43 | + db = Chroma.from_documents(texts, embeddings, persist_directory="./db") |
| 44 | + db.persist() |
| 45 | + retriever = db.as_retriever(search_kwargs={"k": target_source_chunks}) |
| 46 | + return retriever |
| 47 | + except Exception as e: |
| 48 | + logger.error("Error loading documents %s", e, exc_info=True) |
| 49 | + raise e |
| 50 | + |
| 51 | + |
| 52 | +def get_embeddings(): |
| 53 | + return OpenAIEmbeddings() |
| 54 | + |
| 55 | + |
| 56 | +def get_llm(): |
| 57 | + llm = OpenAI(temperature=0.6, model_name="gpt-3.5-turbo-instruct") |
| 58 | + return llm |
| 59 | + |
| 60 | + |
| 61 | +def get_qa_chain(llm, retriever): |
| 62 | + PROMPT = None |
| 63 | + prompt_template = """ |
| 64 | + You are a helpful AI Assistant. |
| 65 | + {summaries} |
| 66 | + Previous Conversations till now: {chat_history} |
| 67 | + Reply to this Human question/instruction: {question}. |
| 68 | + Chatbot: """ |
| 69 | + PROMPT = PromptTemplate( |
| 70 | + template=prompt_template, input_variables=["question", "chat_history"] |
| 71 | + ) |
| 72 | + chain_type_kwargs = {"prompt": PROMPT} |
| 73 | + qa = RetrievalQAWithSourcesChain.from_chain_type( |
| 74 | + llm=llm, |
| 75 | + chain_type="stuff", |
| 76 | + retriever=retriever, |
| 77 | + chain_type_kwargs=chain_type_kwargs, |
| 78 | + memory=ConversationBufferWindowMemory( |
| 79 | + memory_key="chat_history", input_key="question", output_key="answer", k=6 |
| 80 | + ), |
| 81 | + ) |
| 82 | + # qa = LLMChain(prompt=PROMPT, llm=llm, retriever= retriever , memory=ConversationBufferWindowMemory(memory_key="chat_history", input_key="question", k=6), verbose = False) |
| 83 | + return qa |
| 84 | + |
| 85 | + |
| 86 | +def qa_app(qa, query): |
| 87 | + result = qa(query) |
| 88 | + return result["answer"] |
| 89 | + |
| 90 | + |
| 91 | +@app.route("/chatbot/genai/init", methods=["POST"]) |
| 92 | +def init_bot(): |
| 93 | + try: |
| 94 | + with loaded_model_lock: |
| 95 | + if "openai_api_key" in request.json: |
| 96 | + print("Initializing bot", request.json["openai_api_key"]) |
| 97 | + os.environ["OPENAI_API_KEY"] = request.json["openai_api_key"] |
| 98 | + global vulnerable_app_qa, retriever |
| 99 | + retriever = document_loader() |
| 100 | + llm = get_llm() |
| 101 | + vulnerable_app_qa = get_qa_chain(llm, retriever) |
| 102 | + loaded_model = True |
| 103 | + return jsonify({"message": "Model Initialized"}), 200 |
| 104 | + else: |
| 105 | + return jsonify({"message": "openai_api_key not provided"}, 400) |
| 106 | + except Exception as e: |
| 107 | + print("Error initializing bot ", e) |
| 108 | + return jsonify({"message": "Not able to initialize model " + str(e)}), 400 |
| 109 | + |
| 110 | + |
| 111 | +@app.route("/chatbot/genai/state", methods=["GET"]) |
| 112 | +def state_bot(): |
| 113 | + try: |
| 114 | + if loaded_model: |
| 115 | + return jsonify({"message": "Model already loaded"}) |
| 116 | + except Exception as e: |
| 117 | + print("Error checking state ", e) |
| 118 | + return jsonify({"message": "Error checking state " + str(e)}), 400 |
| 119 | + return jsonify({"message": "Model Error"}), 400 |
| 120 | + |
| 121 | + |
| 122 | +@app.route("/chatbot/genai/ask", methods=["POST"]) |
| 123 | +def ask_bot(): |
| 124 | + question = request.json["question"] |
| 125 | + global vulnerable_app_qa |
| 126 | + answer = qa_app(vulnerable_app_qa, question) |
| 127 | + print("###########################################") |
| 128 | + print("Test Attacker Question: " + str(question)) |
| 129 | + print("Vulnerability App Answer: " + str(answer)) |
| 130 | + print("###########################################") |
| 131 | + return jsonify({"answer": answer}), 200 |
| 132 | + |
| 133 | + |
| 134 | +if __name__ == "__main__": |
| 135 | + app.run(host="0.0.0.0", port=5002, debug=True) |
0 commit comments