diff --git a/llm_quantize/quantize.py b/llm_quantize/quantize.py index 7fb1c24..708b6c8 100644 --- a/llm_quantize/quantize.py +++ b/llm_quantize/quantize.py @@ -1,6 +1,6 @@ +import os import subprocess import sys -import os def execute_commands(model_dir_path, quantization=None): @@ -13,7 +13,7 @@ def execute_commands(model_dir_path, quantization=None): if quantization: model_file = f"llama.cpp/models/{model_dir_path}/ggml-model-f16.gguf" quantized_model_file = f"llama.cpp/models/{model_dir_path.split('/')[-1]}/ggml-model-{quantization}.gguf" - subprocess.run(["llama.cpp/llm_quantize", model_file, quantized_model_file, quantization], check=True) + subprocess.run(["llama.cpp/quantize", model_file, quantized_model_file, quantization], check=True) else: print("llama.cpp doesn't exist, check readme how to clone.") diff --git a/projects/Basic-RAG/BasicRAG_stuff.py b/projects/Basic-RAG/BasicRAG_stuff.py index 4bfafc3..63edeab 100644 --- a/projects/Basic-RAG/BasicRAG_stuff.py +++ b/projects/Basic-RAG/BasicRAG_stuff.py @@ -1,6 +1,10 @@ -from grag.grag.rag import BasicRAG +from grag.components.multivec_retriever import Retriever +from grag.components.vectordb.deeplake_client import DeepLakeClient +from grag.rag.basic_rag import BasicRAG -rag = BasicRAG(doc_chain="stuff") +client = DeepLakeClient(collection_name="test") +retriever = Retriever(vectordb=client) +rag = BasicRAG(doc_chain="stuff", retriever=retriever) if __name__ == "__main__": while True: diff --git a/projects/Retriver-GUI/retriever_app.py b/projects/Retriver-GUI/retriever_app.py index f55c0c6..9f4198c 100644 --- a/projects/Retriver-GUI/retriever_app.py +++ b/projects/Retriver-GUI/retriever_app.py @@ -46,7 +46,7 @@ def render_search_results(self): st.write(result.metadata) def check_connection(self): - response = self.app.retriever.client.test_connection() + response = self.app.retriever.vectordb.test_connection() if response: return True else: @@ -55,14 +55,14 @@ def check_connection(self): def render_stats(self): st.write(f''' **Chroma Client Details:** \n - Host Address : {self.app.retriever.client.host}:{self.app.retriever.client.port} \n - Collection Name : {self.app.retriever.client.collection_name} \n - Embeddings Type : {self.app.retriever.client.embedding_type} \n - Embeddings Model: {self.app.retriever.client.embedding_model} \n - Number of docs : {self.app.retriever.client.collection.count()} \n + Host Address : {self.app.retriever.vectordb.host}:{self.app.retriever.vectordb.port} \n + Collection Name : {self.app.retriever.vectordb.collection_name} \n + Embeddings Type : {self.app.retriever.vectordb.embedding_type} \n + Embeddings Model: {self.app.retriever.vectordb.embedding_model} \n + Number of docs : {self.app.retriever.vectordb.collection.count()} \n ''') if st.button('Check Connection'): - response = self.app.retriever.client.test_connection() + response = self.app.retriever.vectordb.test_connection() if response: st.write(':green[Connection Active]') else: diff --git a/pyproject.toml b/pyproject.toml index 3185030..e27f822 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -42,6 +42,7 @@ dependencies = [ "huggingface_hub>=0.20.2", "pydantic>=2.5.0", "rouge-score>=0.1.2", + "deeplake>=3.8.27" ] [project.urls] diff --git a/src/config.ini b/src/config.ini index 54990bf..c2938f9 100644 --- a/src/config.ini +++ b/src/config.ini @@ -14,6 +14,12 @@ n_gpu_layers_cpp : -1 std_out : True base_dir : ${root:root_path}/models +[deeplake] +collection_name : arxiv +embedding_type : instructor-embedding +embedding_model : hkunlp/instructor-xl +store_path : ${data:data_path}/vectordb + [chroma] host : localhost port : 8000 @@ -59,4 +65,7 @@ table_as_html : True data_path : ${root:root_path}/data [root] -root_path : /home/ubuntu/CapStone/Capstone_5 +root_path : /home/ubuntu/volume_2k/Capstone_5 + +[quantize] +llama_cpp_path : ${root:root_path} diff --git a/src/grag/components/multivec_retriever.py b/src/grag/components/multivec_retriever.py index 97684dd..9fd8664 100644 --- a/src/grag/components/multivec_retriever.py +++ b/src/grag/components/multivec_retriever.py @@ -6,11 +6,12 @@ import asyncio import uuid -from typing import List +from typing import Any, Dict, List, Optional -from grag.components.chroma_client import ChromaClient from grag.components.text_splitter import TextSplitter from grag.components.utils import get_config +from grag.components.vectordb.base import VectorDB +from grag.components.vectordb.deeplake_client import DeepLakeClient from langchain.retrievers.multi_vector import MultiVectorRetriever from langchain.storage import LocalFileStore from langchain_core.documents import Document @@ -28,7 +29,7 @@ class Retriever: Attributes: store_path: Path to the local file store id_key: A key prefix for identifying documents - client: ChromaClient class instance from components.chroma_client + vectordb: ChromaClient class instance from components.client store: langchain.storage.LocalFileStore object, stores the key value pairs of document id and parent file retriever: langchain.retrievers.multi_vector.MultiVectorRetriever class instance, langchain's multi-vector retriever splitter: TextSplitter class instance from components.text_splitter @@ -39,10 +40,12 @@ class Retriever: def __init__( self, + vectordb: Optional[VectorDB] = None, store_path: str = multivec_retriever_conf["store_path"], id_key: str = multivec_retriever_conf["id_key"], namespace: str = multivec_retriever_conf["namespace"], top_k=1, + client_kwargs: Optional[Dict[str, Any]] = None, ): """Initialize the Retriever. @@ -55,10 +58,16 @@ def __init__( self.store_path = store_path self.id_key = id_key self.namespace = uuid.UUID(namespace) - self.client = ChromaClient() + if vectordb is None: + if client_kwargs is not None: + self.vectordb = DeepLakeClient(**client_kwargs) + else: + self.vectordb = DeepLakeClient() + else: + self.vectordb = vectordb self.store = LocalFileStore(self.store_path) self.retriever = MultiVectorRetriever( - vectorstore=self.client.langchain_chroma, + vectorstore=self.vectordb.langchain_client, byte_store=self.store, id_key=self.id_key, ) @@ -125,7 +134,7 @@ def add_docs(self, docs: List[Document]): """ chunks = self.split_docs(docs) doc_ids = self.gen_doc_ids(docs) - self.client.add_docs(chunks) + self.vectordb.add_docs(chunks) self.retriever.docstore.mset(list(zip(doc_ids, docs))) async def aadd_docs(self, docs: List[Document]): @@ -140,11 +149,11 @@ async def aadd_docs(self, docs: List[Document]): """ chunks = self.split_docs(docs) doc_ids = self.gen_doc_ids(docs) - await asyncio.run(self.client.aadd_docs(chunks)) + await asyncio.run(self.vectordb.aadd_docs(chunks)) self.retriever.docstore.mset(list(zip(doc_ids))) def get_chunk(self, query: str, with_score=False, top_k=None): - """Returns the most (cosine) similar chunks from the vector database. + """Returns the most similar chunks from the vector database. Args: query: A query string @@ -155,14 +164,8 @@ def get_chunk(self, query: str, with_score=False, top_k=None): list of Documents """ - if with_score: - return self.client.langchain_chroma.similarity_search_with_relevance_scores( - query=query, **{"k": top_k} if top_k else self.retriever.search_kwargs - ) - else: - return self.client.langchain_chroma.similarity_search( - query=query, **{"k": top_k} if top_k else self.retriever.search_kwargs - ) + _top_k = top_k if top_k else self.retriever.search_kwargs["k"] + return self.vectordb.get_chunk(query=query, top_k=_top_k, with_score=with_score) async def aget_chunk(self, query: str, with_score=False, top_k=None): """Returns the most (cosine) similar chunks from the vector database, asynchronously. @@ -176,14 +179,10 @@ async def aget_chunk(self, query: str, with_score=False, top_k=None): list of Documents """ - if with_score: - return await self.client.langchain_chroma.asimilarity_search_with_relevance_scores( - query=query, **{"k": top_k} if top_k else self.retriever.search_kwargs - ) - else: - return await self.client.langchain_chroma.asimilarity_search( - query=query, **{"k": top_k} if top_k else self.retriever.search_kwargs - ) + _top_k = top_k if top_k else self.retriever.search_kwargs["k"] + return await self.vectordb.aget_chunk( + query=query, top_k=_top_k, with_score=with_score + ) def get_doc(self, query: str): """Returns the parent document of the most (cosine) similar chunk from the vector database. diff --git a/src/grag/components/vectordb/__init__.py b/src/grag/components/vectordb/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/src/grag/components/vectordb/base.py b/src/grag/components/vectordb/base.py new file mode 100644 index 0000000..b0b0623 --- /dev/null +++ b/src/grag/components/vectordb/base.py @@ -0,0 +1,85 @@ +"""Abstract base class for vector database clients. + +This module provides: +- VectorDB +""" + +from abc import ABC, abstractmethod +from typing import List, Tuple, Union + +from langchain_community.vectorstores.utils import filter_complex_metadata +from langchain_core.documents import Document + + +class VectorDB(ABC): + """Abstract base class for vector database clients.""" + + @abstractmethod + def __len__(self) -> int: + """Number of chunks in the vector database.""" + ... + + @abstractmethod + def delete(self) -> None: + """Delete all chunks in the vector database.""" + + @abstractmethod + def add_docs(self, docs: List[Document], verbose: bool = True) -> None: + """Adds documents to the vector database. + + Args: + docs: List of Documents + verbose: Show progress bar + + Returns: + None + """ + ... + + @abstractmethod + async def aadd_docs(self, docs: List[Document], verbose: bool = True) -> None: + """Adds documents to the vector database (asynchronous). + + Args: + docs: List of Documents + verbose: Show progress bar + + Returns: + None + """ + ... + + @abstractmethod + def get_chunk( + self, query: str, with_score: bool = False, top_k: int = None + ) -> Union[List[Document], List[Tuple[Document, float]]]: + """Returns the most similar chunks from the vector database. + + Args: + query: A query string + with_score: Outputs scores of returned chunks + top_k: Number of top similar chunks to return, if None defaults to self.top_k + + Returns: + list of Documents + """ + ... + + @abstractmethod + async def aget_chunk( + self, query: str, with_score: bool = False, top_k: int = None + ) -> Union[List[Document], List[Tuple[Document, float]]]: + """Returns the most similar chunks from the vector database (asynchronous). + + Args: + query: A query string + with_score: Outputs scores of returned chunks + top_k: Number of top similar chunks to return, if None defaults to self.top_k + + Returns: + list of Documents + """ + ... + + def _filter_metadata(self, docs: List[Document]) -> List[Document]: + return filter_complex_metadata(docs, allowed_types=self.allowed_metadata_types) diff --git a/src/grag/components/chroma_client.py b/src/grag/components/vectordb/chroma_client.py similarity index 53% rename from src/grag/components/chroma_client.py rename to src/grag/components/vectordb/chroma_client.py index 7efd7c3..cac8ab3 100644 --- a/src/grag/components/chroma_client.py +++ b/src/grag/components/vectordb/chroma_client.py @@ -1,10 +1,16 @@ -from typing import List +"""Class for Chroma vector database. + +This module provides: +- ChromaClient +""" + +from typing import List, Tuple, Union import chromadb from grag.components.embedding import Embedding from grag.components.utils import get_config +from grag.components.vectordb.base import VectorDB from langchain_community.vectorstores import Chroma -from langchain_community.vectorstores.utils import filter_complex_metadata from langchain_core.documents import Document from tqdm import tqdm from tqdm.asyncio import tqdm as atqdm @@ -12,7 +18,7 @@ chroma_conf = get_config()["chroma"] -class ChromaClient: +class ChromaClient(VectorDB): """A class for connecting to a hosted Chroma Vectorstore collection. Attributes: @@ -24,15 +30,15 @@ class ChromaClient: name of the collection in the Chroma Vectorstore, each ChromaClient connects to a single collection embedding_type : str type of embedding used, supported 'sentence-transformers' and 'instructor-embedding' - embedding_modelname : str + embedding_model : str model name of embedding used, should correspond to the embedding_type embedding_function a function of the embedding model, derived from the embedding_type and embedding_modelname - chroma_client + client: chromadb.HttpClient Chroma API for client collection Chroma API for the collection - langchain_chroma + langchain_client: langchain_community.vectorstores.Chroma LangChain wrapper for Chroma collection """ @@ -44,7 +50,9 @@ def __init__( embedding_type=chroma_conf["embedding_type"], embedding_model=chroma_conf["embedding_model"], ): - """Args: + """Initialize a ChromaClient object. + + Args: host: IP Address of hosted Chroma Vectorstore, defaults to argument from config file port: port address of hosted Chroma Vectorstore, defaults to argument from config file collection_name: name of the collection in the Chroma Vectorstore, defaults to argument from config file @@ -61,19 +69,35 @@ def __init__( embedding_model=self.embedding_model, embedding_type=self.embedding_type ).embedding_function - self.chroma_client = chromadb.HttpClient(host=self.host, port=self.port) - self.collection = self.chroma_client.get_or_create_collection( + self.client = chromadb.HttpClient(host=self.host, port=self.port) + self.collection = self.client.get_or_create_collection( name=self.collection_name ) - self.langchain_chroma = Chroma( - client=self.chroma_client, + self.langchain_client = Chroma( + client=self.client, collection_name=self.collection_name, embedding_function=self.embedding_function, ) self.allowed_metadata_types = (str, int, float, bool) - def test_connection(self, verbose=True): - """Tests connection with Chroma Vectorstore + def __len__(self) -> int: + """Count the number of chunks in the database.""" + return self.collection.count() + + def delete(self) -> None: + """Delete all the chunks in the database collection.""" + self.client.delete_collection(self.collection_name) + self.collection = self.client.get_or_create_collection( + name=self.collection_name + ) + self.langchain_client = Chroma( + client=self.client, + collection_name=self.collection_name, + embedding_function=self.embedding_function, + ) + + def test_connection(self, verbose=True) -> int: + """Tests connection with Chroma Vectorstore. Args: verbose: if True, prints connection status @@ -81,7 +105,7 @@ def test_connection(self, verbose=True): Returns: A random integer if connection is alive else None """ - response = self.chroma_client.heartbeat() + response = self.client.heartbeat() if verbose: if response: print(f"Connection to {self.host}/{self.port} is alive..") @@ -89,8 +113,24 @@ def test_connection(self, verbose=True): print(f"Connection to {self.host}/{self.port} is not alive !!") return response - async def aadd_docs(self, docs: List[Document], verbose=True): - """Asynchronously adds documents to chroma vectorstore + def add_docs(self, docs: List[Document], verbose=True) -> None: + """Adds documents to chroma vectorstore. + + Args: + docs: List of Documents + verbose: Show progress bar + + Returns: + None + """ + docs = self._filter_metadata(docs) + for doc in ( + tqdm(docs, desc=f"Adding to {self.collection_name}:") if verbose else docs + ): + _id = self.langchain_client.add_documents([doc]) + + async def aadd_docs(self, docs: List[Document], verbose=True) -> None: + """Asynchronously adds documents to chroma vectorstore. Args: docs: List of Documents @@ -100,37 +140,59 @@ async def aadd_docs(self, docs: List[Document], verbose=True): None """ docs = self._filter_metadata(docs) - # tasks = [self.langchain_chroma.aadd_documents([doc]) for doc in docs] - # if verbose: - # await tqdm_asyncio.gather(*tasks, desc=f'Adding to {self.collection_name}') - # else: - # await asyncio.gather(*tasks) if verbose: for doc in atqdm( docs, desc=f"Adding documents to {self.collection_name}", total=len(docs), ): - await self.langchain_chroma.aadd_documents([doc]) + await self.langchain_client.aadd_documents([doc]) else: for doc in docs: - await self.langchain_chroma.aadd_documents([doc]) + await self.langchain_client.aadd_documents([doc]) - def add_docs(self, docs: List[Document], verbose=True): - """Adds documents to chroma vectorstore + def get_chunk( + self, query: str, with_score: bool = False, top_k: int = None + ) -> Union[List[Document], List[Tuple[Document, float]]]: + """Returns the most similar chunks from the chroma database. Args: - docs: List of Documents - verbose: Show progress bar + query: A query string + with_score: Outputs scores of returned chunks + top_k: Number of top similar chunks to return, if None defaults to self.top_k Returns: - None + list of Documents + """ - docs = self._filter_metadata(docs) - for doc in ( - tqdm(docs, desc=f"Adding to {self.collection_name}:") if verbose else docs - ): - _id = self.langchain_chroma.add_documents([doc]) + if with_score: + return self.langchain_client.similarity_search_with_relevance_scores( + query=query, k=top_k if top_k else 1 + ) + else: + return self.langchain_client.similarity_search( + query=query, k=top_k if top_k else 1 + ) + + async def aget_chunk( + self, query: str, with_score=False, top_k=None + ) -> Union[List[Document], List[Tuple[Document, float]]]: + """Returns the most (cosine) similar chunks from the vector database, asynchronously. + + Args: + query: A query string + with_score: Outputs scores of returned chunks + top_k: Number of top similar chunks to return, if None defaults to self.top_k + + Returns: + list of Documents - def _filter_metadata(self, docs: List[Document]): - return filter_complex_metadata(docs, allowed_types=self.allowed_metadata_types) + """ + if with_score: + return await self.langchain_client.asimilarity_search_with_relevance_scores( + query=query, k=top_k if top_k else 1 + ) + else: + return await self.langchain_client.asimilarity_search( + query=query, k=top_k if top_k else 1 + ) diff --git a/src/grag/components/vectordb/deeplake_client.py b/src/grag/components/vectordb/deeplake_client.py new file mode 100644 index 0000000..f0d5ba5 --- /dev/null +++ b/src/grag/components/vectordb/deeplake_client.py @@ -0,0 +1,159 @@ +"""Class for DeepLake vector database. + +This module provides: +- DeepLakeClient +""" + +from pathlib import Path +from typing import List, Tuple, Union + +from grag.components.embedding import Embedding +from grag.components.utils import get_config +from grag.components.vectordb.base import VectorDB +from langchain_community.vectorstores import DeepLake +from langchain_core.documents import Document +from tqdm import tqdm +from tqdm.asyncio import tqdm as atqdm + +deeplake_conf = get_config()["deeplake"] + + +class DeepLakeClient(VectorDB): + """A class for connecting to a DeepLake Vectorstore. + + Attributes: + store_path : str, Path + The path to store the DeepLake vectorstore. + embedding_type : str + type of embedding used, supported 'sentence-transformers' and 'instructor-embedding' + embedding_model : str + model name of embedding used, should correspond to the embedding_type + embedding_function + a function of the embedding model, derived from the embedding_type and embedding_modelname + client: deeplake.core.vectorstore.VectorStore + DeepLake API + collection + Chroma API for the collection + langchain_client: langchain_community.vectorstores.DeepLake + LangChain wrapper for DeepLake API + """ + + def __init__( + self, + collection_name: str = deeplake_conf["collection_name"], + store_path: Union[str, Path] = deeplake_conf["store_path"], + embedding_type: str = deeplake_conf["embedding_type"], + embedding_model: str = deeplake_conf["embedding_model"], + read_only: bool = False, + ): + """Initialize DeepLake client object.""" + self.store_path = Path(store_path) + self.collection_name = collection_name + self.read_only = read_only + self.embedding_type: str = embedding_type + self.embedding_model: str = embedding_model + + self.embedding_function = Embedding( + embedding_model=self.embedding_model, embedding_type=self.embedding_type + ).embedding_function + + # self.client = VectorStore(path=self.store_path / self.collection_name) + self.langchain_client = DeepLake( + dataset_path=str(self.store_path / self.collection_name), + embedding=self.embedding_function, + read_only=self.read_only, + ) + self.client = self.langchain_client.vectorstore + self.allowed_metadata_types = (str, int, float, bool) + + def __len__(self) -> int: + """Number of chunks in the vector database.""" + return self.client.__len__() + + def delete(self) -> None: + """Delete all chunks in the vector database.""" + self.client.delete(delete_all=True) + + def add_docs(self, docs: List[Document], verbose=True) -> None: + """Adds documents to deeplake vectorstore. + + Args: + docs: List of Documents + verbose: Show progress bar + + Returns: + None + """ + docs = self._filter_metadata(docs) + for doc in ( + tqdm(docs, desc=f"Adding to {self.collection_name}:") if verbose else docs + ): + _id = self.langchain_client.add_documents([doc]) + + async def aadd_docs(self, docs: List[Document], verbose=True) -> None: + """Asynchronously adds documents to chroma vectorstore. + + Args: + docs: List of Documents + verbose: Show progress bar + + Returns: + None + """ + docs = self._filter_metadata(docs) + if verbose: + for doc in atqdm( + docs, + desc=f"Adding documents to {self.collection_name}", + total=len(docs), + ): + await self.langchain_client.aadd_documents([doc]) + else: + for doc in docs: + await self.langchain_client.aadd_documents([doc]) + + def get_chunk( + self, query: str, with_score: bool = False, top_k: int = None + ) -> Union[List[Document], List[Tuple[Document, float]]]: + """Returns the most similar chunks from the deeplake database. + + Args: + query: A query string + with_score: Outputs scores of returned chunks + top_k: Number of top similar chunks to return, if None defaults to self.top_k + + Returns: + list of Documents + + """ + if with_score: + return self.langchain_client.similarity_search_with_score( + query=query, k=top_k if top_k else 1 + ) + else: + return self.langchain_client.similarity_search( + query=query, k=top_k if top_k else 1 + ) + + async def aget_chunk( + self, query: str, with_score=False, top_k=None + ) -> Union[List[Document], List[Tuple[Document, float]]]: + """Returns the most similar chunks from the deeplake database, asynchronously. + + Args: + query: A query string + with_score: Outputs scores of returned chunks + top_k: Number of top similar chunks to return, if None defaults to self.top_k + + Returns: + list of Documents + + """ + if with_score: + return await self.langchain_client.asimilarity_search_with_score( + query=query, k=top_k if top_k else 1 + ) + else: + return await self.langchain_client.asimilarity_search( + query=query, k=top_k if top_k else 1 + ) diff --git a/src/grag/quantize/__init__.py b/src/grag/quantize/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/src/grag/quantize/quantize.py b/src/grag/quantize/quantize.py new file mode 100644 index 0000000..64fba47 --- /dev/null +++ b/src/grag/quantize/quantize.py @@ -0,0 +1,52 @@ +"""Interactive file for quantizing models.""" + +from pathlib import Path + +from grag.components.utils import get_config +from grag.quantize.utils import ( + building_llamacpp, + fetch_model_repo, + get_llamacpp_repo, + quantize_model, +) + +config = get_config() +root_path = Path(config["quantize"]["llama_cpp_path"]) + +if __name__ == "__main__": + user_input = input( + "Enter the path to the llama_cpp cloned repo, or where you'd like to clone it. Press Enter to use the default config path: " + ).strip() + + if user_input != "": + root_path = Path(user_input) + + res = get_llamacpp_repo(root_path) + + if "Already up to date." in str(res.stdout): + print("Repository is already up to date. Skipping build.") + else: + print("Updates found. Starting build...") + building_llamacpp(root_path) + + response = ( + input("Do you want us to download the model? (y/n) [Enter for yes]: ") + .strip() + .lower() + ) + if response == "n": + print("Please copy the model folder to 'llama.cpp/models/' folder.") + _ = input("Enter if you have already copied the model:") + model_dir = Path(input("Enter the model directory name: ")) + elif response == "y" or response == "": + repo_id = input( + "Please enter the repo_id for the model (you can check on https://huggingface.co/models): " + ).strip() + fetch_model_repo(repo_id, root_path) + # model_dir = repo_id.split('/')[1] + model_dir = root_path / "llama.cpp" / "models" / repo_id.split("/")[1] + + quantization = input( + "Enter quantization, recommended - Q5_K_M or Q4_K_M for more check https://github.com/ggerganov/llama.cpp/blob/master/examples/quantize/quantize.cpp#L19 : " + ) + quantize_model(model_dir, quantization, root_path) diff --git a/src/grag/quantize/utils.py b/src/grag/quantize/utils.py new file mode 100644 index 0000000..bc1d280 --- /dev/null +++ b/src/grag/quantize/utils.py @@ -0,0 +1,135 @@ +"""Utility functions for quantization.""" + +import os +import subprocess +from pathlib import Path +from typing import Optional, Union + +from grag.components.utils import get_config +from huggingface_hub import snapshot_download + +config = get_config() + + +def get_llamacpp_repo(root_path: Union[str, Path]) -> subprocess.CompletedProcess: + """Clones or pulls the llama.cpp repository into the specified root path. + + Args: + root_path: The root directory where the llama.cpp repository will be cloned or updated. + + Returns: + A subprocess.CompletedProcess instance containing the result of the git operation. + """ + if os.path.exists(f"{root_path}/llama.cpp"): + print(f"Repo exists at: {root_path}/llama.cpp") + res = subprocess.run( + ["git", "-C", f"{root_path}/llama.cpp", "pull"], + check=True, + capture_output=True, + ) + else: + res = subprocess.run( + [ + "git", + "clone", + "https://github.com/ggerganov/llama.cpp.git", + f"{root_path}/llama.cpp", + ], + check=True, + capture_output=True, + ) + + return res + + +def building_llamacpp(root_path: Union[str, Path]) -> None: + """Attempts to build the llama.cpp project using make or cmake. + + Args: + root_path (str): The root directory where the llama.cpp project is located. + """ + os.chdir(f"{root_path}/llama.cpp/") + try: + subprocess.run(["which", "make"], check=True, stdout=subprocess.DEVNULL) + subprocess.run(["make", "LLAMA_CUBLAS=1"], check=True) + print("Llama.cpp build successful.") + except subprocess.CalledProcessError: + try: + subprocess.run(["which", "cmake"], check=True, stdout=subprocess.DEVNULL) + subprocess.run(["mkdir", "build"], check=True) + subprocess.run( + [ + "cd", + "build", + "&&", + "cmake", + "..", + "-DLLAMA_CUBLAS=ON", + "&&", + "cmake", + "--build", + ".", + "--config", + "Release", + ], + shell=True, + check=True, + ) + print("Llama.cpp build successful.") + except subprocess.CalledProcessError: + print("Unable to build, cannot find make or cmake.") + finally: + os.chdir( + Path(__file__).parent + ) # Assuming you want to return to the root path after operation + + +def fetch_model_repo(repo_id: str, root_path: Union[str, Path]) -> None: + """Download model from huggingface.co/models. + + Args: + repo_id (str): Repository ID of the model to download. + root_path (str): The root path where the model should be downloaded or copied. + """ + local_dir = f"{root_path}/llama.cpp/models/{repo_id.split('/')[1]}" + os.makedirs(local_dir, exist_ok=True) + snapshot_download( + repo_id=repo_id, + local_dir=local_dir, + local_dir_use_symlinks="auto", + resume_download=True, + ) + print(f"Model downloaded in {local_dir}") + + +def quantize_model( + model_dir_path: Union[str, Path], + quantization: str, + root_path: Union[str, Path], + output_dir: Optional[Union[str, Path]] = None, +) -> None: + """Quantizes a specified model using a given quantization level. + + Args: + output_dir (str, Path, optional): Directory to save quantized model. Defaults to None + model_dir_path (str, Path): The directory path of the model to be quantized. + quantization (str): The quantization level to apply. + root_path (str, Path): The root directory path of the project. + """ + os.chdir(f"{root_path}/llama.cpp/") + model_dir_path = Path(model_dir_path) + if output_dir is None: + output_dir = config["llm"]["base_dir"] + + output_dir = Path(output_dir) / model_dir_path.name + os.makedirs(output_dir, exist_ok=True) + + subprocess.run(["python3", "convert.py", f"{model_dir_path}/"], check=True) + model_file = model_dir_path / "ggml-model-f32.gguf" + quantized_model_file = output_dir / f"ggml-model-{quantization}.gguf" + subprocess.run( + ["./quantize", str(model_file), str(quantized_model_file), quantization], + check=True, + ) + print(f"Quantized model present at {output_dir}") + os.chdir(Path(__file__).parent) # Return to the root path after operation diff --git a/src/grag/rag/basic_rag.py b/src/grag/rag/basic_rag.py index be055ba..da461b6 100644 --- a/src/grag/rag/basic_rag.py +++ b/src/grag/rag/basic_rag.py @@ -5,7 +5,7 @@ """ import json -from typing import List, Union +from typing import List, Optional, Union from grag import prompts from grag.components.llm import LLM @@ -32,20 +32,21 @@ class BasicRAG: def __init__( self, + retriever: Optional[Retriever] = None, model_name=None, doc_chain="stuff", task="QA", llm_kwargs=None, retriever_kwargs=None, - custom_prompt: Union[ - Prompt, FewShotPrompt, List[Prompt, FewShotPrompt], None - ] = None, + custom_prompt: Union[Prompt, FewShotPrompt, None] = None, ): - """Initialize BasicRAG.""" - if retriever_kwargs is None: - self.retriever = Retriever() + if retriever is None: + if retriever_kwargs is None: + self.retriever = Retriever() + else: + self.retriever = Retriever(**retriever_kwargs) else: - self.retriever = Retriever(**retriever_kwargs) + self.retriever = retriever if llm_kwargs is None: self.llm_ = LLM() diff --git a/src/tests/components/multivec_retriever_test.py b/src/tests/components/multivec_retriever_test.py index 3ccb3fb..14dad0b 100644 --- a/src/tests/components/multivec_retriever_test.py +++ b/src/tests/components/multivec_retriever_test.py @@ -1,3 +1,92 @@ +import json + +from grag.components.multivec_retriever import Retriever +from langchain_core.documents import Document + +retriever = Retriever() # pass test collection + +doc = Document(page_content="Hello worlds", metadata={"source": "bars"}) + + +def test_retriver_id_gen(): + doc = Document(page_content="Hello world", metadata={"source": "bar"}) + id_ = retriever.id_gen(doc) + assert isinstance(id, str) + assert len(id_) == 32 + doc.page_content = doc.page_content + 'ABC' + id_1 = retriever.id_gen(doc) + assert id_ == id_1 + doc.metadata["source"] = "bars" + id_1 = retriever.id_gen(doc) + assert id_ != id_1 + + +def test_retriever_gen_doc_ids(): + docs = [Document(page_content="Hello world", metadata={"source": "bar"}), + Document(page_content="Hello", metadata={"source": "foo"})] + ids = retriever.gen_doc_ids(docs) + assert len(ids) == len(docs) + assert all(isinstance(id, str) for id in ids) + + +def test_retriever_split_docs(): + pass + + +def test_retriever_split_docs(): + pass + + +def test_retriever_add_docs(): + # small enough docs to not split. + docs = [Document(page_content= + """And so on this rainbow day, with storms all around them, and blue sky + above, they rode only as far as the valley. But from there, before they + turned to go back, the monuments appeared close, and they loomed + grandly with the background of purple bank and creamy cloud and shafts + of golden lightning. They seemed like sentinels--guardians of a great + and beautiful love born under their lofty heights, in the lonely + silence of day, in the star-thrown shadow of night. They were like that + love. And they held Lucy and Slone, calling every day, giving a + nameless and tranquil content, binding them true to love, true to the + sage and the open, true to that wild upland home""", metadata={"source": "test_doc_1"}), + Document(page_content= + """Slone and Lucy never rode down so far as the stately monuments, though + these held memories as hauntingly sweet as others were poignantly + bitter. Lucy never rode the King again. But Slone rode him, learned to + love him. And Lucy did not race any more. When Slone tried to stir in + her the old spirit all the response he got was a wistful shake of head + or a laugh that hid the truth or an excuse that the strain on her + ankles from Joel Creech's lasso had never mended. The girl was + unutterably happy, but it was possible that she would never race a + horse again.""", metadata={"source": "test_doc_2"}), + Document(page_content= + """Bostil wanted to be alone, to welcome the King, to lead him back to the + home corral, perhaps to hide from all eyes the change and the uplift + that would forever keep him from wronging another man. + + The late rains came and like magic, in a few days, the sage grew green + and lustrous and fresh, the gray turning to purple. + + Every morning the sun rose white and hot in a blue and cloudless sky. + And then soon the horizon line showed creamy clouds that rose and + spread and darkened. Every afternoon storms hung along the ramparts and + rainbows curved down beautiful and ethereal. The dim blackness of the + storm-clouds was split to the blinding zigzag of lightning, and the + thunder rolled and boomed, like the Colorado in flood.""", metadata={"source": "test_doc_3"}) + ] + ids = retriever.gen_doc_ids(docs) + retriever.add_docs(docs) + retrieved = retriever.store.mget(ids) + assert len(retrieved) == len(ids) + for i, doc in enumerate(docs): + retrieved_doc = json.loads(retrieved[i].decode()) + assert doc.metadata == retrieved_doc.metadata + + +def test_retriever_aadd_docs(): + pass + # # add code folder to sys path # import os # from pathlib import Path diff --git a/src/tests/components/vectordb/__init__.py b/src/tests/components/vectordb/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/src/tests/components/chroma_client_test.py b/src/tests/components/vectordb/chroma_client_test.py similarity index 59% rename from src/tests/components/chroma_client_test.py rename to src/tests/components/vectordb/chroma_client_test.py index 1596dd3..c491dfd 100644 --- a/src/tests/components/chroma_client_test.py +++ b/src/tests/components/vectordb/chroma_client_test.py @@ -1,12 +1,13 @@ import asyncio -from grag.components.chroma_client import ChromaClient +import pytest +from grag.components.vectordb.chroma_client import ChromaClient from langchain_core.documents import Document def test_chroma_connection(): - client = ChromaClient() - response = client.test_connection() + chroma_client = ChromaClient() + response = chroma_client.test_connection() assert isinstance(response, int) @@ -45,14 +46,12 @@ def test_chroma_add_docs(): storm-clouds was split to the blinding zigzag of lightning, and the thunder rolled and boomed, like the Colorado in flood.""", ] - client = ChromaClient(collection_name="test") - if client.collection.count() > 0: - client.chroma_client.delete_collection("test") - client = ChromaClient(collection_name="test") + chroma_client = ChromaClient(collection_name="test") + if len(chroma_client) > 0: + chroma_client.delete() docs = [Document(page_content=doc) for doc in docs] - client.add_docs(docs) - collection_count = client.collection.count() - assert collection_count == len(docs) + chroma_client.add_docs(docs) + assert len(chroma_client) == len(docs) def test_chroma_aadd_docs(): @@ -90,11 +89,60 @@ def test_chroma_aadd_docs(): storm-clouds was split to the blinding zigzag of lightning, and the thunder rolled and boomed, like the Colorado in flood.""", ] - client = ChromaClient(collection_name="test") - if client.collection.count() > 0: - client.chroma_client.delete_collection("test") - client = ChromaClient(collection_name="test") + chroma_client = ChromaClient(collection_name="test") + if len(chroma_client) > 0: + chroma_client.delete() docs = [Document(page_content=doc) for doc in docs] loop = asyncio.get_event_loop() - loop.run_until_complete(client.aadd_docs(docs)) - assert client.collection.count() == len(docs) + loop.run_until_complete(chroma_client.aadd_docs(docs)) + assert len(chroma_client) == len(docs) + + +chrome_get_chunk_params = [(1, False), (1, True), (2, False), (2, True)] + + +@pytest.mark.parametrize("top_k,with_score", chrome_get_chunk_params) +def test_chroma_get_chunk(top_k, with_score): + query = """Slone and Lucy never rode down so far as the stately monuments, though + these held memories as hauntingly sweet as others were poignantly + bitter. Lucy never rode the King again. But Slone rode him, learned to + love him. And Lucy did not race any more. When Slone tried to stir in + her the old spirit all the response he got was a wistful shake of head + or a laugh that hid the truth or an excuse that the strain on her + ankles from Joel Creech's lasso had never mended. The girl was + unutterably happy, but it was possible that she would never race a + horse again.""" + chroma_client = ChromaClient(collection_name="test") + retrieved_chunks = chroma_client.get_chunk( + query=query, top_k=top_k, with_score=with_score + ) + assert len(retrieved_chunks) == top_k + if with_score: + assert all(isinstance(doc[0], Document) for doc in retrieved_chunks) + assert all(isinstance(doc[1], float) for doc in retrieved_chunks) + else: + assert all(isinstance(doc, Document) for doc in retrieved_chunks) + + +@pytest.mark.parametrize("top_k,with_score", chrome_get_chunk_params) +def test_chroma_aget_chunk(top_k, with_score): + query = """Slone and Lucy never rode down so far as the stately monuments, though + these held memories as hauntingly sweet as others were poignantly + bitter. Lucy never rode the King again. But Slone rode him, learned to + love him. And Lucy did not race any more. When Slone tried to stir in + her the old spirit all the response he got was a wistful shake of head + or a laugh that hid the truth or an excuse that the strain on her + ankles from Joel Creech's lasso had never mended. The girl was + unutterably happy, but it was possible that she would never race a + horse again.""" + chroma_client = ChromaClient(collection_name="test") + loop = asyncio.get_event_loop() + retrieved_chunks = loop.run_until_complete( + chroma_client.aget_chunk(query=query, top_k=top_k, with_score=with_score) + ) + assert len(retrieved_chunks) == top_k + if with_score: + assert all(isinstance(doc[0], Document) for doc in retrieved_chunks) + assert all(isinstance(doc[1], float) for doc in retrieved_chunks) + else: + assert all(isinstance(doc, Document) for doc in retrieved_chunks) diff --git a/src/tests/components/vectordb/deeplake_client_test.py b/src/tests/components/vectordb/deeplake_client_test.py new file mode 100644 index 0000000..cea5e61 --- /dev/null +++ b/src/tests/components/vectordb/deeplake_client_test.py @@ -0,0 +1,146 @@ +import asyncio + +import pytest +from grag.components.vectordb.deeplake_client import DeepLakeClient +from langchain_core.documents import Document + + +def test_deeplake_add_docs(): + docs = [ + """And so on this rainbow day, with storms all around them, and blue sky + above, they rode only as far as the valley. But from there, before they + turned to go back, the monuments appeared close, and they loomed + grandly with the background of purple bank and creamy cloud and shafts + of golden lightning. They seemed like sentinels--guardians of a great + and beautiful love born under their lofty heights, in the lonely + silence of day, in the star-thrown shadow of night. They were like that + love. And they held Lucy and Slone, calling every day, giving a + nameless and tranquil content, binding them true to love, true to the + sage and the open, true to that wild upland home.""", + """Slone and Lucy never rode down so far as the stately monuments, though + these held memories as hauntingly sweet as others were poignantly + bitter. Lucy never rode the King again. But Slone rode him, learned to + love him. And Lucy did not race any more. When Slone tried to stir in + her the old spirit all the response he got was a wistful shake of head + or a laugh that hid the truth or an excuse that the strain on her + ankles from Joel Creech's lasso had never mended. The girl was + unutterably happy, but it was possible that she would never race a + horse again.""", + """Bostil wanted to be alone, to welcome the King, to lead him back to the + home corral, perhaps to hide from all eyes the change and the uplift + that would forever keep him from wronging another man. + + The late rains came and like magic, in a few days, the sage grew green + and lustrous and fresh, the gray turning to purple. + + Every morning the sun rose white and hot in a blue and cloudless sky. + And then soon the horizon line showed creamy clouds that rose and + spread and darkened. Every afternoon storms hung along the ramparts and + rainbows curved down beautiful and ethereal. The dim blackness of the + storm-clouds was split to the blinding zigzag of lightning, and the + thunder rolled and boomed, like the Colorado in flood.""", + ] + deeplake_client = DeepLakeClient(collection_name="test") + if len(deeplake_client) > 0: + deeplake_client.delete() + docs = [Document(page_content=doc) for doc in docs] + deeplake_client.add_docs(docs) + assert len(deeplake_client) == len(docs) + del deeplake_client + + +def test_chroma_aadd_docs(): + docs = [ + """And so on this rainbow day, with storms all around them, and blue sky + above, they rode only as far as the valley. But from there, before they + turned to go back, the monuments appeared close, and they loomed + grandly with the background of purple bank and creamy cloud and shafts + of golden lightning. They seemed like sentinels--guardians of a great + and beautiful love born under their lofty heights, in the lonely + silence of day, in the star-thrown shadow of night. They were like that + love. And they held Lucy and Slone, calling every day, giving a + nameless and tranquil content, binding them true to love, true to the + sage and the open, true to that wild upland home.""", + """Slone and Lucy never rode down so far as the stately monuments, though + these held memories as hauntingly sweet as others were poignantly + bitter. Lucy never rode the King again. But Slone rode him, learned to + love him. And Lucy did not race any more. When Slone tried to stir in + her the old spirit all the response he got was a wistful shake of head + or a laugh that hid the truth or an excuse that the strain on her + ankles from Joel Creech's lasso had never mended. The girl was + unutterably happy, but it was possible that she would never race a + horse again.""", + """Bostil wanted to be alone, to welcome the King, to lead him back to the + home corral, perhaps to hide from all eyes the change and the uplift + that would forever keep him from wronging another man. + + The late rains came and like magic, in a few days, the sage grew green + and lustrous and fresh, the gray turning to purple. + + Every morning the sun rose white and hot in a blue and cloudless sky. + And then soon the horizon line showed creamy clouds that rose and + spread and darkened. Every afternoon storms hung along the ramparts and + rainbows curved down beautiful and ethereal. The dim blackness of the + storm-clouds was split to the blinding zigzag of lightning, and the + thunder rolled and boomed, like the Colorado in flood.""", + ] + deeplake_client = DeepLakeClient(collection_name="test") + if len(deeplake_client) > 0: + deeplake_client.delete() + docs = [Document(page_content=doc) for doc in docs] + loop = asyncio.get_event_loop() + loop.run_until_complete(deeplake_client.aadd_docs(docs)) + assert len(deeplake_client) == len(docs) + del deeplake_client + + +deeplake_get_chunk_params = [(1, False), (1, True), (2, False), (2, True)] + + +@pytest.mark.parametrize("top_k,with_score", deeplake_get_chunk_params) +def test_deeplake_get_chunk(top_k, with_score): + query = """Slone and Lucy never rode down so far as the stately monuments, though + these held memories as hauntingly sweet as others were poignantly + bitter. Lucy never rode the King again. But Slone rode him, learned to + love him. And Lucy did not race any more. When Slone tried to stir in + her the old spirit all the response he got was a wistful shake of head + or a laugh that hid the truth or an excuse that the strain on her + ankles from Joel Creech's lasso had never mended. The girl was + unutterably happy, but it was possible that she would never race a + horse again.""" + deeplake_client = DeepLakeClient(collection_name="test", read_only=True) + retrieved_chunks = deeplake_client.get_chunk( + query=query, top_k=top_k, with_score=with_score + ) + assert len(retrieved_chunks) == top_k + if with_score: + assert all(isinstance(doc[0], Document) for doc in retrieved_chunks) + assert all(isinstance(doc[1], float) for doc in retrieved_chunks) + else: + assert all(isinstance(doc, Document) for doc in retrieved_chunks) + del deeplake_client + + +@pytest.mark.parametrize("top_k,with_score", deeplake_get_chunk_params) +def test_deeplake_aget_chunk(top_k, with_score): + query = """Slone and Lucy never rode down so far as the stately monuments, though + these held memories as hauntingly sweet as others were poignantly + bitter. Lucy never rode the King again. But Slone rode him, learned to + love him. And Lucy did not race any more. When Slone tried to stir in + her the old spirit all the response he got was a wistful shake of head + or a laugh that hid the truth or an excuse that the strain on her + ankles from Joel Creech's lasso had never mended. The girl was + unutterably happy, but it was possible that she would never race a + horse again.""" + deeplake_client = DeepLakeClient(collection_name="test", read_only=True) + loop = asyncio.get_event_loop() + retrieved_chunks = loop.run_until_complete( + deeplake_client.aget_chunk(query=query, top_k=top_k, with_score=with_score) + ) + assert len(retrieved_chunks) == top_k + if with_score: + assert all(isinstance(doc[0], Document) for doc in retrieved_chunks) + assert all(isinstance(doc[1], float) for doc in retrieved_chunks) + else: + assert all(isinstance(doc, Document) for doc in retrieved_chunks) + del deeplake_client diff --git a/src/tests/quantize/__init__.py b/src/tests/quantize/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/src/tests/quantize/quantize_test.py b/src/tests/quantize/quantize_test.py new file mode 100644 index 0000000..af0e9dd --- /dev/null +++ b/src/tests/quantize/quantize_test.py @@ -0,0 +1,39 @@ +import os +from pathlib import Path + +from grag.quantize.utils import ( + building_llamacpp, + fetch_model_repo, + get_llamacpp_repo, + quantize_model, +) + +root_path = Path(__file__).parent / "test_data" +os.makedirs(root_path, exist_ok=True) + + +def test_get_llamacpp_repo(): + get_llamacpp_repo(root_path) + repo_path = root_path / "llama.cpp" / ".git" + assert os.path.exists(repo_path) + + +def test_build_llamacpp(): + building_llamacpp(root_path) + bin_path = root_path / "llama.cpp" / "quantize" + assert os.path.exists(bin_path) + + +def test_fetch_model_repo(): + fetch_model_repo("meta-llama/Llama-2-7b-chat", root_path) + model_dir_path = root_path / "llama.cpp" / "models" / "Llama-2-7b-chat" + assert os.path.exists(model_dir_path) + + +def test_quantize_model(): + model_dir_path = root_path / "llama.cpp" / "models" / "Llama-2-7b-chat" + quantize_model( + model_dir_path, "Q3_K_M", root_path, output_dir=model_dir_path.parent + ) + gguf_file_path = model_dir_path / "ggml-model-Q3_K_M.gguf" + assert os.path.exists(gguf_file_path) diff --git a/src/tests/rag/basic_rag_test.py b/src/tests/rag/basic_rag_test.py index 06db25e..2249028 100644 --- a/src/tests/rag/basic_rag_test.py +++ b/src/tests/rag/basic_rag_test.py @@ -1,4 +1,4 @@ -from typing import List, Text +from typing import Text, List from grag.rag.basic_rag import BasicRAG