import importlib.metadata import logging import os import threading import uuid from typing import Optional import requests from dotenv import load_dotenv from langchain.docstore.document import Document from langchain.memory import ConversationBufferMemory from tenacity import retry, stop_after_attempt, wait_fixed from embedchain.chunkers.base_chunker import BaseChunker from embedchain.config import AddConfig, ChatConfig, QueryConfig from embedchain.config.apps.BaseAppConfig import BaseAppConfig from embedchain.config.QueryConfig import DOCS_SITE_PROMPT_TEMPLATE from embedchain.data_formatter import DataFormatter from embedchain.loaders.base_loader import BaseLoader load_dotenv() ABS_PATH = os.getcwd() DB_DIR = os.path.join(ABS_PATH, "db") memory = ConversationBufferMemory() class EmbedChain: def __init__(self, config: BaseAppConfig): """ Initializes the EmbedChain instance, sets up a vector DB client and creates a collection. :param config: BaseAppConfig instance to load as configuration. """ self.config = config self.collection = self.config.db._get_or_create_collection(self.config.collection_name) self.db = self.config.db self.user_asks = [] self.is_docs_site_instance = False self.online = False # Send anonymous telemetry self.s_id = self.config.id if self.config.id else str(uuid.uuid4()) thread_telemetry = threading.Thread(target=self._send_telemetry_event, args=("init",)) thread_telemetry.start() def add(self, data_type, url, metadata=None, config: AddConfig = None): """ Adds the data from the given URL to the vector db. Loads the data, chunks it, create embedding for each chunk and then stores the embedding to vector database. :param data_type: The type of the data to add. :param url: The URL where the data is located. :param metadata: Optional. Metadata associated with the data source. :param config: Optional. The `AddConfig` instance to use as configuration options. """ if config is None: config = AddConfig() data_formatter = DataFormatter(data_type, config) self.user_asks.append([data_type, url, metadata]) documents, _metadatas, _ids, new_chunks = self.load_and_embed( data_formatter.loader, data_formatter.chunker, url, metadata ) if data_type in ("docs_site",): self.is_docs_site_instance = True # Send anonymous telemetry if self.config.collect_metrics: # it's quicker to check the variable twice than to count words when they won't be submitted. word_count = sum([len(document.split(" ")) for document in documents]) extra_metadata = {"data_type": data_type, "word_count": word_count, "chunks_count": new_chunks} thread_telemetry = threading.Thread(target=self._send_telemetry_event, args=("add", extra_metadata)) thread_telemetry.start() def add_local(self, data_type, content, metadata=None, config: AddConfig = None): """ Adds the data you supply to the vector db. Loads the data, chunks it, create embedding for each chunk and then stores the embedding to vector database. :param data_type: The type of the data to add. :param content: The local data. Refer to the `README` for formatting. :param metadata: Optional. Metadata associated with the data source. :param config: Optional. The `AddConfig` instance to use as configuration options. """ if config is None: config = AddConfig() data_formatter = DataFormatter(data_type, config) self.user_asks.append([data_type, content]) documents, _metadatas, _ids, new_chunks = self.load_and_embed( data_formatter.loader, data_formatter.chunker, content, metadata ) # Send anonymous telemetry if self.config.collect_metrics: # it's quicker to check the variable twice than to count words when they won't be submitted. word_count = sum([len(document.split(" ")) for document in documents]) extra_metadata = {"data_type": data_type, "word_count": word_count, "chunks_count": new_chunks} thread_telemetry = threading.Thread(target=self._send_telemetry_event, args=("add_local", extra_metadata)) thread_telemetry.start() def load_and_embed(self, loader: BaseLoader, chunker: BaseChunker, src, metadata=None): """ Loads the data from the given URL, chunks it, and adds it to database. :param loader: The loader to use to load the data. :param chunker: The chunker to use to chunk the data. :param src: The data to be handled by the loader. Can be a URL for remote sources or local content for local loaders. :param metadata: Optional. Metadata associated with the data source. :return: (List) documents (embedded text), (List) metadata, (list) ids, (int) number of chunks """ embeddings_data = chunker.create_chunks(loader, src) documents = embeddings_data["documents"] metadatas = embeddings_data["metadatas"] ids = embeddings_data["ids"] # get existing ids, and discard doc if any common id exist. where = {"app_id": self.config.id} if self.config.id is not None else {} # where={"url": src} existing_ids = self.db.get( ids=ids, where=where, # optional filter ) if len(existing_ids): data_dict = {id: (doc, meta) for id, doc, meta in zip(ids, documents, metadatas)} data_dict = {id: value for id, value in data_dict.items() if id not in existing_ids} if not data_dict: print(f"All data from {src} already exists in the database.") # Make sure to return a matching return type return [], [], [], 0 ids = list(data_dict.keys()) documents, metadatas = zip(*data_dict.values()) # Add app id in metadatas so that they can be queried on later if self.config.id is not None: metadatas = [{**m, "app_id": self.config.id} for m in metadatas] # FIXME: Fix the error handling logic when metadatas or metadata is None metadatas = metadatas if metadatas else [] metadata = metadata if metadata else {} chunks_before_addition = self.count() # Add metadata to each document metadatas_with_metadata = [{**meta, **metadata} for meta in metadatas] self.db.add(documents=documents, metadatas=metadatas_with_metadata, ids=ids) count_new_chunks = self.count() - chunks_before_addition print((f"Successfully saved {src}. New chunks count: {count_new_chunks}")) return list(documents), metadatas_with_metadata, ids, count_new_chunks def _format_result(self, results): return [ (Document(page_content=result[0], metadata=result[1] or {}), result[2]) for result in zip( results["documents"][0], results["metadatas"][0], results["distances"][0], ) ] def get_llm_model_answer(self): """ Usually implemented by child class """ raise NotImplementedError def retrieve_from_database(self, input_query, config: QueryConfig): """ Queries the vector database based on the given input query. Gets relevant doc based on the query :param input_query: The query to use. :param config: The query configuration. :return: The content of the document that matched your query. """ where = {"app_id": self.config.id} if self.config.id is not None else {} # optional filter contents = self.db.query( input_query=input_query, n_results=config.number_documents, where=where, ) return contents def _append_search_and_context(self, context, web_search_result): return f"{context}\nWeb Search Result: {web_search_result}" def generate_prompt(self, input_query, contexts, config: QueryConfig, **kwargs): """ Generates a prompt based on the given query and context, ready to be passed to an LLM :param input_query: The query to use. :param contexts: List of similar documents to the query used as context. :param config: Optional. The `QueryConfig` instance to use as configuration options. :return: The prompt """ context_string = (" | ").join(contexts) web_search_result = kwargs.get("web_search_result", "") if web_search_result: context_string = self._append_search_and_context(context_string, web_search_result) if not config.history: prompt = config.template.substitute(context=context_string, query=input_query) else: prompt = config.template.substitute(context=context_string, query=input_query, history=config.history) return prompt def get_answer_from_llm(self, prompt, config: ChatConfig): """ Gets an answer based on the given query and context by passing it to an LLM. :param query: The query to use. :param context: Similar documents to the query used as context. :return: The answer. """ return self.get_llm_model_answer(prompt, config) def access_search_and_get_results(self, input_query): from langchain.tools import DuckDuckGoSearchRun search = DuckDuckGoSearchRun() logging.info(f"Access search to get answers for {input_query}") return search.run(input_query) def query(self, input_query, config: QueryConfig = None, dry_run=False): """ Queries the vector database based on the given input query. Gets relevant doc based on the query and then passes it to an LLM as context to get the answer. :param input_query: The query to use. :param config: Optional. The `QueryConfig` instance to use as configuration options. :param dry_run: Optional. A dry run does everything except send the resulting prompt to the LLM. The purpose is to test the prompt, not the response. You can use it to test your prompt, including the context provided by the vector database's doc retrieval. The only thing the dry run does not consider is the cut-off due to the `max_tokens` parameter. :return: The answer to the query. """ if config is None: config = QueryConfig() if self.is_docs_site_instance: config.template = DOCS_SITE_PROMPT_TEMPLATE config.number_documents = 5 k = {} if self.online: k["web_search_result"] = self.access_search_and_get_results(input_query) contexts = self.retrieve_from_database(input_query, config) prompt = self.generate_prompt(input_query, contexts, config, **k) logging.info(f"Prompt: {prompt}") if dry_run: return prompt answer = self.get_answer_from_llm(prompt, config) # Send anonymous telemetry thread_telemetry = threading.Thread(target=self._send_telemetry_event, args=("query",)) thread_telemetry.start() if isinstance(answer, str): logging.info(f"Answer: {answer}") return answer else: return self._stream_query_response(answer) def _stream_query_response(self, answer): streamed_answer = "" for chunk in answer: streamed_answer = streamed_answer + chunk yield chunk logging.info(f"Answer: {streamed_answer}") def chat(self, input_query, config: ChatConfig = None, dry_run=False): """ Queries the vector database on the given input query. Gets relevant doc based on the query and then passes it to an LLM as context to get the answer. Maintains the whole conversation in memory. :param input_query: The query to use. :param config: Optional. The `ChatConfig` instance to use as configuration options. :param dry_run: Optional. A dry run does everything except send the resulting prompt to the LLM. The purpose is to test the prompt, not the response. You can use it to test your prompt, including the context provided by the vector database's doc retrieval. The only thing the dry run does not consider is the cut-off due to the `max_tokens` parameter. :return: The answer to the query. """ if config is None: config = ChatConfig() if self.is_docs_site_instance: config.template = DOCS_SITE_PROMPT_TEMPLATE config.number_documents = 5 k = {} if self.online: k["web_search_result"] = self.access_search_and_get_results(input_query) contexts = self.retrieve_from_database(input_query, config) global memory chat_history = memory.load_memory_variables({})["history"] if chat_history: config.set_history(chat_history) prompt = self.generate_prompt(input_query, contexts, config, **k) logging.info(f"Prompt: {prompt}") if dry_run: return prompt answer = self.get_answer_from_llm(prompt, config) memory.chat_memory.add_user_message(input_query) # Send anonymous telemetry thread_telemetry = threading.Thread(target=self._send_telemetry_event, args=("chat",)) thread_telemetry.start() if isinstance(answer, str): memory.chat_memory.add_ai_message(answer) logging.info(f"Answer: {answer}") return answer else: # this is a streamed response and needs to be handled differently. return self._stream_chat_response(answer) def _stream_chat_response(self, answer): streamed_answer = "" for chunk in answer: streamed_answer = streamed_answer + chunk yield chunk memory.chat_memory.add_ai_message(streamed_answer) logging.info(f"Answer: {streamed_answer}") def set_collection(self, collection_name): """ Set the collection to use. :param collection_name: The name of the collection to use. """ self.collection = self.config.db._get_or_create_collection(collection_name) def count(self) -> int: """ Count the number of embeddings. :return: The number of embeddings. """ return self.db.count() def reset(self): """ Resets the database. Deletes all embeddings irreversibly. `App` does not have to be reinitialized after using this method. """ # Send anonymous telemetry thread_telemetry = threading.Thread(target=self._send_telemetry_event, args=("reset",)) thread_telemetry.start() collection_name = self.collection.name self.db.reset() self.collection = self.config.db._get_or_create_collection(collection_name) # Todo: Automatically recreating a collection with the same name cannot be the best way to handle a reset. # A downside of this implementation is, if you have two instances, # the other instance will not get the updated `self.collection` attribute. # A better way would be to create the collection if it is called again after being reset. # That means, checking if collection exists in the db-consuming methods, and creating it if it doesn't. # That's an extra steps for all uses, just to satisfy a niche use case in a niche method. For now, this will do. @retry(stop=stop_after_attempt(3), wait=wait_fixed(1)) def _send_telemetry_event(self, method: str, extra_metadata: Optional[dict] = None): if not self.config.collect_metrics: return with threading.Lock(): url = "https://api.embedchain.ai/api/v1/telemetry/" metadata = { "s_id": self.s_id, "version": importlib.metadata.version(__package__ or __name__), "method": method, "language": "py", } if extra_metadata: metadata.update(extra_metadata) response = requests.post(url, json={"metadata": metadata}) response.raise_for_status()