# Copyright (C) 2024 Intel Corporation # SPDX-License-Identifier: Apache-2.0 import argparse import json import os import re from comps import MegaServiceEndpoint, MicroService, ServiceOrchestrator, ServiceRoleType, ServiceType from comps.cores.mega.utils import handle_message from comps.cores.proto.api_protocol import ( ChatCompletionRequest, ChatCompletionResponse, ChatCompletionResponseChoice, ChatMessage, UsageInfo, ) from comps.cores.proto.docarray import LLMParams, RerankerParms, RetrieverParms from fastapi import Request from fastapi.responses import StreamingResponse from langchain_core.prompts import PromptTemplate class ChatTemplate: @staticmethod def generate_rag_prompt(question, documents): context_str = "\n".join(documents) if context_str and len(re.findall("[\u4E00-\u9FFF]", context_str)) / len(context_str) >= 0.3: # chinese context template = """ ### 你将扮演一个乐于助人、尊重他人并诚实的助手，你的目标是帮助用户解答问题。有效地利用来自本地知识库的搜索结果。确保你的回答中只包含相关信息。如果你不确定问题的答案，请避免分享不准确的信息。 ### 搜索结果：{context} ### 问题：{question} ### 回答： """ else: template = """ ### You are a helpful, respectful and honest assistant to help the user with questions. \ Please refer to the search results obtained from the local knowledge base. \ But be careful to not incorporate the information that you think is not relevant to the question. \ If you don't know the answer to a question, please don't share false information. \n ### Search results: {context} \n ### Question: {question} \n ### Answer: """ return template.format(context=context_str, question=question) MEGA_SERVICE_PORT = int(os.getenv("MEGA_SERVICE_PORT", 8888)) GUARDRAIL_SERVICE_HOST_IP = os.getenv("GUARDRAIL_SERVICE_HOST_IP", "0.0.0.0") GUARDRAIL_SERVICE_PORT = int(os.getenv("GUARDRAIL_SERVICE_PORT", 80)) EMBEDDING_SERVER_HOST_IP = os.getenv("EMBEDDING_SERVER_HOST_IP", "0.0.0.0") EMBEDDING_SERVER_PORT = int(os.getenv("EMBEDDING_SERVER_PORT", 80)) RETRIEVER_SERVICE_HOST_IP = os.getenv("RETRIEVER_SERVICE_HOST_IP", "0.0.0.0") RETRIEVER_SERVICE_PORT = int(os.getenv("RETRIEVER_SERVICE_PORT", 7000)) RERANK_SERVER_HOST_IP = os.getenv("RERANK_SERVER_HOST_IP", "0.0.0.0") RERANK_SERVER_PORT = int(os.getenv("RERANK_SERVER_PORT", 80)) LLM_SERVER_HOST_IP = os.getenv("LLM_SERVER_HOST_IP", "0.0.0.0") LLM_SERVER_PORT = int(os.getenv("LLM_SERVER_PORT", 80)) LLM_MODEL = os.getenv("LLM_MODEL", "meta-llama/Meta-Llama-3-8B-Instruct") def align_inputs(self, inputs, cur_node, runtime_graph, llm_parameters_dict, **kwargs): if self.services[cur_node].service_type == ServiceType.EMBEDDING: inputs["inputs"] = inputs["text"] del inputs["text"] elif self.services[cur_node].service_type == ServiceType.RETRIEVER: # prepare the retriever params retriever_parameters = kwargs.get("retriever_parameters", None) if retriever_parameters: inputs.update(retriever_parameters.dict()) elif self.services[cur_node].service_type == ServiceType.LLM: # convert TGI/vLLM to unified OpenAI /v1/chat/completions format next_inputs = {} next_inputs["model"] = LLM_MODEL next_inputs["messages"] = [{"role": "user", "content": inputs["inputs"]}] next_inputs["max_tokens"] = llm_parameters_dict["max_tokens"] next_inputs["top_p"] = llm_parameters_dict["top_p"] next_inputs["stream"] = inputs["stream"] next_inputs["frequency_penalty"] = inputs["frequency_penalty"] # next_inputs["presence_penalty"] = inputs["presence_penalty"] # next_inputs["repetition_penalty"] = inputs["repetition_penalty"] next_inputs["temperature"] = inputs["temperature"] inputs = next_inputs return inputs def align_outputs(self, data, cur_node, inputs, runtime_graph, llm_parameters_dict, **kwargs): next_data = {} if self.services[cur_node].service_type == ServiceType.EMBEDDING: assert isinstance(data, list) next_data = {"text": inputs["inputs"], "embedding": data[0]} elif self.services[cur_node].service_type == ServiceType.RETRIEVER: docs = [doc["text"] for doc in data["retrieved_docs"]] with_rerank = runtime_graph.downstream(cur_node)[0].startswith("rerank") if with_rerank and docs: # forward to rerank # prepare inputs for rerank next_data["query"] = data["initial_query"] next_data["texts"] = [doc["text"] for doc in data["retrieved_docs"]] else: # forward to llm if not docs and with_rerank: # delete the rerank from retriever -> rerank -> llm for ds in reversed(runtime_graph.downstream(cur_node)): for nds in runtime_graph.downstream(ds): runtime_graph.add_edge(cur_node, nds) runtime_graph.delete_node_if_exists(ds) # handle template # if user provides template, then format the prompt with it # otherwise, use the default template prompt = data["initial_query"] chat_template = llm_parameters_dict["chat_template"] if chat_template: prompt_template = PromptTemplate.from_template(chat_template) input_variables = prompt_template.input_variables if sorted(input_variables) == ["context", "question"]: prompt = prompt_template.format(question=data["initial_query"], context="\n".join(docs)) elif input_variables == ["question"]: prompt = prompt_template.format(question=data["initial_query"]) else: print(f"{prompt_template} not used, we only support 2 input variables ['question', 'context']") prompt = ChatTemplate.generate_rag_prompt(data["initial_query"], docs) else: prompt = ChatTemplate.generate_rag_prompt(data["initial_query"], docs) next_data["inputs"] = prompt elif self.services[cur_node].service_type == ServiceType.RERANK: # rerank the inputs with the scores reranker_parameters = kwargs.get("reranker_parameters", None) top_n = reranker_parameters.top_n if reranker_parameters else 1 docs = inputs["texts"] reranked_docs = [] for best_response in data[:top_n]: reranked_docs.append(docs[best_response["index"]]) # handle template # if user provides template, then format the prompt with it # otherwise, use the default template prompt = inputs["query"] chat_template = llm_parameters_dict["chat_template"] if chat_template: prompt_template = PromptTemplate.from_template(chat_template) input_variables = prompt_template.input_variables if sorted(input_variables) == ["context", "question"]: prompt = prompt_template.format(question=prompt, context="\n".join(reranked_docs)) elif input_variables == ["question"]: prompt = prompt_template.format(question=prompt) else: print(f"{prompt_template} not used, we only support 2 input variables ['question', 'context']") prompt = ChatTemplate.generate_rag_prompt(prompt, reranked_docs) else: prompt = ChatTemplate.generate_rag_prompt(prompt, reranked_docs) next_data["inputs"] = prompt elif self.services[cur_node].service_type == ServiceType.LLM and not llm_parameters_dict["stream"]: next_data["text"] = data["choices"][0]["message"]["content"] else: next_data = data return next_data def align_generator(self, gen, **kwargs): # OpenAI response format # b'data:{"id":"","object":"text_completion","created":1725530204,"model":"meta-llama/Meta-Llama-3-8B-Instruct","system_fingerprint":"2.0.1-native","choices":[{"index":0,"delta":{"role":"assistant","content":"?"},"logprobs":null,"finish_reason":null}]}\n\n' for line in gen: line = line.decode("utf-8") start = line.find("{") end = line.rfind("}") + 1 json_str = line[start:end] try: # sometimes yield empty chunk, do a fallback here json_data = json.loads(json_str) if ( json_data["choices"][0]["finish_reason"] != "eos_token" and "content" in json_data["choices"][0]["delta"] ): yield f"data: {repr(json_data['choices'][0]['delta']['content'].encode('utf-8'))}\n\n" except Exception as e: yield f"data: {repr(json_str.encode('utf-8'))}\n\n" yield "data: [DONE]\n\n" class ChatQnAService: def __init__(self, host="0.0.0.0", port=8000): self.host = host self.port = port ServiceOrchestrator.align_inputs = align_inputs ServiceOrchestrator.align_outputs = align_outputs ServiceOrchestrator.align_generator = align_generator self.megaservice = ServiceOrchestrator() self.endpoint = str(MegaServiceEndpoint.CHAT_QNA) def add_remote_service(self): embedding = MicroService( name="embedding", host=EMBEDDING_SERVER_HOST_IP, port=EMBEDDING_SERVER_PORT, endpoint="/embed", use_remote_service=True, service_type=ServiceType.EMBEDDING, ) retriever = MicroService( name="retriever", host=RETRIEVER_SERVICE_HOST_IP, port=RETRIEVER_SERVICE_PORT, endpoint="/v1/retrieval", use_remote_service=True, service_type=ServiceType.RETRIEVER, ) rerank = MicroService( name="rerank", host=RERANK_SERVER_HOST_IP, port=RERANK_SERVER_PORT, endpoint="/rerank", use_remote_service=True, service_type=ServiceType.RERANK, ) llm = MicroService( name="llm", host=LLM_SERVER_HOST_IP, port=LLM_SERVER_PORT, endpoint="/v1/chat/completions", use_remote_service=True, service_type=ServiceType.LLM, ) self.megaservice.add(embedding).add(retriever).add(rerank).add(llm) self.megaservice.flow_to(embedding, retriever) self.megaservice.flow_to(retriever, rerank) self.megaservice.flow_to(rerank, llm) def add_remote_service_without_rerank(self): embedding = MicroService( name="embedding", host=EMBEDDING_SERVER_HOST_IP, port=EMBEDDING_SERVER_PORT, endpoint="/embed", use_remote_service=True, service_type=ServiceType.EMBEDDING, ) retriever = MicroService( name="retriever", host=RETRIEVER_SERVICE_HOST_IP, port=RETRIEVER_SERVICE_PORT, endpoint="/v1/retrieval", use_remote_service=True, service_type=ServiceType.RETRIEVER, ) llm = MicroService( name="llm", host=LLM_SERVER_HOST_IP, port=LLM_SERVER_PORT, endpoint="/v1/chat/completions", use_remote_service=True, service_type=ServiceType.LLM, ) self.megaservice.add(embedding).add(retriever).add(llm) self.megaservice.flow_to(embedding, retriever) self.megaservice.flow_to(retriever, llm) def add_remote_service_with_guardrails(self): guardrail_in = MicroService( name="guardrail_in", host=GUARDRAIL_SERVICE_HOST_IP, port=GUARDRAIL_SERVICE_PORT, endpoint="/v1/guardrails", use_remote_service=True, service_type=ServiceType.GUARDRAIL, ) embedding = MicroService( name="embedding", host=EMBEDDING_SERVER_HOST_IP, port=EMBEDDING_SERVER_PORT, endpoint="/embed", use_remote_service=True, service_type=ServiceType.EMBEDDING, ) retriever = MicroService( name="retriever", host=RETRIEVER_SERVICE_HOST_IP, port=RETRIEVER_SERVICE_PORT, endpoint="/v1/retrieval", use_remote_service=True, service_type=ServiceType.RETRIEVER, ) rerank = MicroService( name="rerank", host=RERANK_SERVER_HOST_IP, port=RERANK_SERVER_PORT, endpoint="/rerank", use_remote_service=True, service_type=ServiceType.RERANK, ) llm = MicroService( name="llm", host=LLM_SERVER_HOST_IP, port=LLM_SERVER_PORT, endpoint="/v1/chat/completions", use_remote_service=True, service_type=ServiceType.LLM, ) # guardrail_out = MicroService( # name="guardrail_out", # host=GUARDRAIL_SERVICE_HOST_IP, # port=GUARDRAIL_SERVICE_PORT, # endpoint="/v1/guardrails", # use_remote_service=True, # service_type=ServiceType.GUARDRAIL, # ) # self.megaservice.add(guardrail_in).add(embedding).add(retriever).add(rerank).add(llm).add(guardrail_out) self.megaservice.add(guardrail_in).add(embedding).add(retriever).add(rerank).add(llm) self.megaservice.flow_to(guardrail_in, embedding) self.megaservice.flow_to(embedding, retriever) self.megaservice.flow_to(retriever, rerank) self.megaservice.flow_to(rerank, llm) # self.megaservice.flow_to(llm, guardrail_out) async def handle_request(self, request: Request): data = await request.json() stream_opt = data.get("stream", True) chat_request = ChatCompletionRequest.parse_obj(data) prompt = handle_message(chat_request.messages) parameters = LLMParams( max_tokens=chat_request.max_tokens if chat_request.max_tokens else 1024, top_k=chat_request.top_k if chat_request.top_k else 10, top_p=chat_request.top_p if chat_request.top_p else 0.95, temperature=chat_request.temperature if chat_request.temperature else 0.01, frequency_penalty=chat_request.frequency_penalty if chat_request.frequency_penalty else 0.0, presence_penalty=chat_request.presence_penalty if chat_request.presence_penalty else 0.0, repetition_penalty=chat_request.repetition_penalty if chat_request.repetition_penalty else 1.03, stream=stream_opt, chat_template=chat_request.chat_template if chat_request.chat_template else None, ) retriever_parameters = RetrieverParms( search_type=chat_request.search_type if chat_request.search_type else "similarity", k=chat_request.k if chat_request.k else 4, distance_threshold=chat_request.distance_threshold if chat_request.distance_threshold else None, fetch_k=chat_request.fetch_k if chat_request.fetch_k else 20, lambda_mult=chat_request.lambda_mult if chat_request.lambda_mult else 0.5, score_threshold=chat_request.score_threshold if chat_request.score_threshold else 0.2, ) reranker_parameters = RerankerParms( top_n=chat_request.top_n if chat_request.top_n else 1, ) result_dict, runtime_graph = await self.megaservice.schedule( initial_inputs={"text": prompt}, llm_parameters=parameters, retriever_parameters=retriever_parameters, reranker_parameters=reranker_parameters, ) for node, response in result_dict.items(): if isinstance(response, StreamingResponse): return response last_node = runtime_graph.all_leaves()[-1] response = result_dict[last_node]["text"] choices = [] usage = UsageInfo() choices.append( ChatCompletionResponseChoice( index=0, message=ChatMessage(role="assistant", content=response), finish_reason="stop", ) ) return ChatCompletionResponse(model="chatqna", choices=choices, usage=usage) def start(self): self.service = MicroService( self.__class__.__name__, service_role=ServiceRoleType.MEGASERVICE, host=self.host, port=self.port, endpoint=self.endpoint, input_datatype=ChatCompletionRequest, output_datatype=ChatCompletionResponse, ) self.service.add_route(self.endpoint, self.handle_request, methods=["POST"]) self.service.start() if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument("--without-rerank", action="store_true") parser.add_argument("--with-guardrails", action="store_true") args = parser.parse_args() chatqna = ChatQnAService(port=MEGA_SERVICE_PORT) if args.without_rerank: chatqna.add_remote_service_without_rerank() elif args.with_guardrails: chatqna.add_remote_service_with_guardrails() else: chatqna.add_remote_service() chatqna.start()