diff --git a/fern/docs.yml b/fern/docs.yml index 2611dac89..be0d904c2 100644 --- a/fern/docs.yml +++ b/fern/docs.yml @@ -64,6 +64,8 @@ navigation: contents: - page: LLM Backends path: ./docs/pages/manual/llms.mdx + - page: Reranking + path: ./docs/pages/manual/reranker.mdx - section: User Interface contents: - page: User interface (Gradio) Manual diff --git a/fern/docs/pages/manual/reranker.mdx b/fern/docs/pages/manual/reranker.mdx new file mode 100644 index 000000000..576e3e8f0 --- /dev/null +++ b/fern/docs/pages/manual/reranker.mdx @@ -0,0 +1,36 @@ +## Enhancing Response Quality with Reranking + +PrivateGPT offers a reranking feature aimed at optimizing response generation by filtering out irrelevant documents, potentially leading to faster response times and enhanced relevance of answers generated by the LLM. + +### Enabling Reranking + +Document reranking can significantly improve the efficiency and quality of the responses by pre-selecting the most relevant documents before generating an answer. To leverage this feature, ensure that it is enabled in the RAG settings and consider adjusting the parameters to best fit your use case. + +#### Additional Requirements + +Before enabling reranking, you must install additional dependencies: + +```bash +poetry install --extras rerank-sentence-transformers +``` + +This command installs dependencies for the cross-encoder reranker from sentence-transformers, which is currently the only supported method by PrivateGPT for document reranking. + +#### Configuration + +To enable and configure reranking, adjust the `rag` section within the `settings.yaml` file. Here are the key settings to consider: + +- `similarity_top_k`: Determines the number of documents to initially retrieve and consider for reranking. This value should be larger than `top_n`. +- `rerank`: + - `enabled`: Set to `true` to activate the reranking feature. + - `top_n`: Specifies the number of documents to use in the final answer generation process, chosen from the top-ranked documents provided by `similarity_top_k`. + +Example configuration snippet: + +```yaml +rag: + similarity_top_k: 10 # Number of documents to retrieve and consider for reranking + rerank: + enabled: true + top_n: 3 # Number of top-ranked documents to use for generating the answer +``` \ No newline at end of file diff --git a/poetry.lock b/poetry.lock index 38c2374c0..e7f186dda 100644 --- a/poetry.lock +++ b/poetry.lock @@ -4949,6 +4949,90 @@ tensorflow = ["safetensors[numpy]", "tensorflow (>=2.11.0)"] testing = ["h5py (>=3.7.0)", "huggingface_hub (>=0.12.1)", "hypothesis (>=6.70.2)", "pytest (>=7.2.0)", "pytest-benchmark (>=4.0.0)", "safetensors[numpy]", "setuptools_rust (>=1.5.2)"] torch = ["safetensors[numpy]", "torch (>=1.10)"] +[[package]] +name = "scikit-learn" +version = "1.4.1.post1" +description = "A set of python modules for machine learning and data mining" +optional = true +python-versions = ">=3.9" +files = [ + {file = "scikit-learn-1.4.1.post1.tar.gz", hash = "sha256:93d3d496ff1965470f9977d05e5ec3376fb1e63b10e4fda5e39d23c2d8969a30"}, + {file = "scikit_learn-1.4.1.post1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:c540aaf44729ab5cd4bd5e394f2b375e65ceaea9cdd8c195788e70433d91bbc5"}, + {file = "scikit_learn-1.4.1.post1-cp310-cp310-macosx_12_0_arm64.whl", hash = "sha256:4310bff71aa98b45b46cd26fa641309deb73a5d1c0461d181587ad4f30ea3c36"}, + {file = "scikit_learn-1.4.1.post1-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:9f43dd527dabff5521af2786a2f8de5ba381e182ec7292663508901cf6ceaf6e"}, + {file = "scikit_learn-1.4.1.post1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:c02e27d65b0c7dc32f2c5eb601aaf5530b7a02bfbe92438188624524878336f2"}, + {file = "scikit_learn-1.4.1.post1-cp310-cp310-win_amd64.whl", hash = "sha256:629e09f772ad42f657ca60a1a52342eef786218dd20cf1369a3b8d085e55ef8f"}, + {file = "scikit_learn-1.4.1.post1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:6145dfd9605b0b50ae72cdf72b61a2acd87501369a763b0d73d004710ebb76b5"}, + {file = "scikit_learn-1.4.1.post1-cp311-cp311-macosx_12_0_arm64.whl", hash = "sha256:1afed6951bc9d2053c6ee9a518a466cbc9b07c6a3f9d43bfe734192b6125d508"}, + {file = "scikit_learn-1.4.1.post1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ce03506ccf5f96b7e9030fea7eb148999b254c44c10182ac55857bc9b5d4815f"}, + {file = "scikit_learn-1.4.1.post1-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:4ba516fcdc73d60e7f48cbb0bccb9acbdb21807de3651531208aac73c758e3ab"}, + {file = "scikit_learn-1.4.1.post1-cp311-cp311-win_amd64.whl", hash = "sha256:78cd27b4669513b50db4f683ef41ea35b5dddc797bd2bbd990d49897fd1c8a46"}, + {file = "scikit_learn-1.4.1.post1-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:a1e289f33f613cefe6707dead50db31930530dc386b6ccff176c786335a7b01c"}, + {file = "scikit_learn-1.4.1.post1-cp312-cp312-macosx_12_0_arm64.whl", hash = "sha256:0df87de9ce1c0140f2818beef310fb2e2afdc1e66fc9ad587965577f17733649"}, + {file = "scikit_learn-1.4.1.post1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:712c1c69c45b58ef21635360b3d0a680ff7d83ac95b6f9b82cf9294070cda710"}, + {file = "scikit_learn-1.4.1.post1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:1754b0c2409d6ed5a3380512d0adcf182a01363c669033a2b55cca429ed86a81"}, + {file = "scikit_learn-1.4.1.post1-cp312-cp312-win_amd64.whl", hash = "sha256:1d491ef66e37f4e812db7e6c8286520c2c3fc61b34bf5e59b67b4ce528de93af"}, + {file = "scikit_learn-1.4.1.post1-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:aa0029b78ef59af22cfbd833e8ace8526e4df90212db7ceccbea582ebb5d6794"}, + {file = "scikit_learn-1.4.1.post1-cp39-cp39-macosx_12_0_arm64.whl", hash = "sha256:14e4c88436ac96bf69eb6d746ac76a574c314a23c6961b7d344b38877f20fee1"}, + {file = "scikit_learn-1.4.1.post1-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d7cd3a77c32879311f2aa93466d3c288c955ef71d191503cf0677c3340ae8ae0"}, + {file = "scikit_learn-1.4.1.post1-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:2a3ee19211ded1a52ee37b0a7b373a8bfc66f95353af058a210b692bd4cda0dd"}, + {file = "scikit_learn-1.4.1.post1-cp39-cp39-win_amd64.whl", hash = "sha256:234b6bda70fdcae9e4abbbe028582ce99c280458665a155eed0b820599377d25"}, +] + +[package.dependencies] +joblib = ">=1.2.0" +numpy = ">=1.19.5,<2.0" +scipy = ">=1.6.0" +threadpoolctl = ">=2.0.0" + +[package.extras] +benchmark = ["matplotlib (>=3.3.4)", "memory-profiler (>=0.57.0)", "pandas (>=1.1.5)"] +docs = ["Pillow (>=7.1.2)", "matplotlib (>=3.3.4)", "memory-profiler (>=0.57.0)", "numpydoc (>=1.2.0)", "pandas (>=1.1.5)", "plotly (>=5.14.0)", "pooch (>=1.6.0)", "scikit-image (>=0.17.2)", "seaborn (>=0.9.0)", "sphinx (>=6.0.0)", "sphinx-copybutton (>=0.5.2)", "sphinx-gallery (>=0.15.0)", "sphinx-prompt (>=1.3.0)", "sphinxext-opengraph (>=0.4.2)"] +examples = ["matplotlib (>=3.3.4)", "pandas (>=1.1.5)", "plotly (>=5.14.0)", "pooch (>=1.6.0)", "scikit-image (>=0.17.2)", "seaborn (>=0.9.0)"] +tests = ["black (>=23.3.0)", "matplotlib (>=3.3.4)", "mypy (>=1.3)", "numpydoc (>=1.2.0)", "pandas (>=1.1.5)", "polars (>=0.19.12)", "pooch (>=1.6.0)", "pyamg (>=4.0.0)", "pyarrow (>=12.0.0)", "pytest (>=7.1.2)", "pytest-cov (>=2.9.0)", "ruff (>=0.0.272)", "scikit-image (>=0.17.2)"] + +[[package]] +name = "scipy" +version = "1.12.0" +description = "Fundamental algorithms for scientific computing in Python" +optional = true +python-versions = ">=3.9" +files = [ + {file = "scipy-1.12.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:78e4402e140879387187f7f25d91cc592b3501a2e51dfb320f48dfb73565f10b"}, + {file = "scipy-1.12.0-cp310-cp310-macosx_12_0_arm64.whl", hash = "sha256:f5f00ebaf8de24d14b8449981a2842d404152774c1a1d880c901bf454cb8e2a1"}, + {file = "scipy-1.12.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e53958531a7c695ff66c2e7bb7b79560ffdc562e2051644c5576c39ff8efb563"}, + {file = "scipy-1.12.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:5e32847e08da8d895ce09d108a494d9eb78974cf6de23063f93306a3e419960c"}, + {file = "scipy-1.12.0-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:4c1020cad92772bf44b8e4cdabc1df5d87376cb219742549ef69fc9fd86282dd"}, + {file = "scipy-1.12.0-cp310-cp310-win_amd64.whl", hash = "sha256:75ea2a144096b5e39402e2ff53a36fecfd3b960d786b7efd3c180e29c39e53f2"}, + {file = "scipy-1.12.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:408c68423f9de16cb9e602528be4ce0d6312b05001f3de61fe9ec8b1263cad08"}, + {file = "scipy-1.12.0-cp311-cp311-macosx_12_0_arm64.whl", hash = "sha256:5adfad5dbf0163397beb4aca679187d24aec085343755fcdbdeb32b3679f254c"}, + {file = "scipy-1.12.0-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c3003652496f6e7c387b1cf63f4bb720951cfa18907e998ea551e6de51a04467"}, + {file = "scipy-1.12.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8b8066bce124ee5531d12a74b617d9ac0ea59245246410e19bca549656d9a40a"}, + {file = "scipy-1.12.0-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:8bee4993817e204d761dba10dbab0774ba5a8612e57e81319ea04d84945375ba"}, + {file = "scipy-1.12.0-cp311-cp311-win_amd64.whl", hash = "sha256:a24024d45ce9a675c1fb8494e8e5244efea1c7a09c60beb1eeb80373d0fecc70"}, + {file = "scipy-1.12.0-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:e7e76cc48638228212c747ada851ef355c2bb5e7f939e10952bc504c11f4e372"}, + {file = "scipy-1.12.0-cp312-cp312-macosx_12_0_arm64.whl", hash = "sha256:f7ce148dffcd64ade37b2df9315541f9adad6efcaa86866ee7dd5db0c8f041c3"}, + {file = "scipy-1.12.0-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:9c39f92041f490422924dfdb782527a4abddf4707616e07b021de33467f917bc"}, + {file = "scipy-1.12.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a7ebda398f86e56178c2fa94cad15bf457a218a54a35c2a7b4490b9f9cb2676c"}, + {file = "scipy-1.12.0-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:95e5c750d55cf518c398a8240571b0e0782c2d5a703250872f36eaf737751338"}, + {file = "scipy-1.12.0-cp312-cp312-win_amd64.whl", hash = "sha256:e646d8571804a304e1da01040d21577685ce8e2db08ac58e543eaca063453e1c"}, + {file = "scipy-1.12.0-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:913d6e7956c3a671de3b05ccb66b11bc293f56bfdef040583a7221d9e22a2e35"}, + {file = "scipy-1.12.0-cp39-cp39-macosx_12_0_arm64.whl", hash = "sha256:bba1b0c7256ad75401c73e4b3cf09d1f176e9bd4248f0d3112170fb2ec4db067"}, + {file = "scipy-1.12.0-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:730badef9b827b368f351eacae2e82da414e13cf8bd5051b4bdfd720271a5371"}, + {file = "scipy-1.12.0-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:6546dc2c11a9df6926afcbdd8a3edec28566e4e785b915e849348c6dd9f3f490"}, + {file = "scipy-1.12.0-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:196ebad3a4882081f62a5bf4aeb7326aa34b110e533aab23e4374fcccb0890dc"}, + {file = "scipy-1.12.0-cp39-cp39-win_amd64.whl", hash = "sha256:b360f1b6b2f742781299514e99ff560d1fe9bd1bff2712894b52abe528d1fd1e"}, + {file = "scipy-1.12.0.tar.gz", hash = "sha256:4bf5abab8a36d20193c698b0f1fc282c1d083c94723902c447e5d2f1780936a3"}, +] + +[package.dependencies] +numpy = ">=1.22.4,<1.29.0" + +[package.extras] +dev = ["click", "cython-lint (>=0.12.2)", "doit (>=0.36.0)", "mypy", "pycodestyle", "pydevtool", "rich-click", "ruff", "types-psutil", "typing_extensions"] +doc = ["jupytext", "matplotlib (>2)", "myst-nb", "numpydoc", "pooch", "pydata-sphinx-theme (==0.9.0)", "sphinx (!=4.1.0)", "sphinx-design (>=0.2.0)"] +test = ["asv", "gmpy2", "hypothesis", "mpmath", "pooch", "pytest", "pytest-cov", "pytest-timeout", "pytest-xdist", "scikit-umfpack", "threadpoolctl"] + [[package]] name = "semantic-version" version = "2.10.0" @@ -4964,6 +5048,27 @@ files = [ dev = ["Django (>=1.11)", "check-manifest", "colorama (<=0.4.1)", "coverage", "flake8", "nose2", "readme-renderer (<25.0)", "tox", "wheel", "zest.releaser[recommended]"] doc = ["Sphinx", "sphinx-rtd-theme"] +[[package]] +name = "sentence-transformers" +version = "2.6.1" +description = "Multilingual text embeddings" +optional = true +python-versions = ">=3.8.0" +files = [ + {file = "sentence-transformers-2.6.1.tar.gz", hash = "sha256:633ad6b70e390ea335de8689652a5d6c21a323b79ed19519c2f392451088487f"}, + {file = "sentence_transformers-2.6.1-py3-none-any.whl", hash = "sha256:a887e17696b513f99a709ce1f37fd547f53857aebe863785ede546c303b09ea0"}, +] + +[package.dependencies] +huggingface-hub = ">=0.15.1" +numpy = "*" +Pillow = "*" +scikit-learn = "*" +scipy = "*" +torch = ">=1.11.0" +tqdm = "*" +transformers = ">=4.32.0,<5.0.0" + [[package]] name = "setuptools" version = "69.0.2" @@ -5156,6 +5261,17 @@ files = [ [package.extras] doc = ["reno", "sphinx", "tornado (>=4.5)"] +[[package]] +name = "threadpoolctl" +version = "3.4.0" +description = "threadpoolctl" +optional = true +python-versions = ">=3.8" +files = [ + {file = "threadpoolctl-3.4.0-py3-none-any.whl", hash = "sha256:8f4c689a65b23e5ed825c8436a92b818aac005e0f3715f6a1664d7c7ee29d262"}, + {file = "threadpoolctl-3.4.0.tar.gz", hash = "sha256:f11b491a03661d6dd7ef692dd422ab34185d982466c49c8f98c8f716b5c93196"}, +] + [[package]] name = "tiktoken" version = "0.5.2" @@ -6206,6 +6322,7 @@ llms-ollama = ["llama-index-llms-ollama"] llms-openai = ["llama-index-llms-openai"] llms-openai-like = ["llama-index-llms-openai-like"] llms-sagemaker = ["boto3"] +rerank-sentence-transformers = ["sentence-transformers", "torch"] storage-nodestore-postgres = ["asyncpg", "llama-index-storage-docstore-postgres", "llama-index-storage-index-store-postgres", "psycopg2-binary"] ui = ["gradio"] vector-stores-chroma = ["llama-index-vector-stores-chroma"] @@ -6215,4 +6332,4 @@ vector-stores-qdrant = ["llama-index-vector-stores-qdrant"] [metadata] lock-version = "2.0" python-versions = ">=3.11,<3.12" -content-hash = "3d5f21e5e41ea66d655891a6d9b01bcdd8348b275e27a54e90b65ac9d5719981" +content-hash = "0b3665bd11a604609249ff0267e4e5cf009881d16a84f9774fc54d45a1373e09" diff --git a/private_gpt/server/chat/chat_service.py b/private_gpt/server/chat/chat_service.py index ea57f2c0d..ae8cf0089 100644 --- a/private_gpt/server/chat/chat_service.py +++ b/private_gpt/server/chat/chat_service.py @@ -9,6 +9,7 @@ from llama_index.core.indices.postprocessor import MetadataReplacementPostProcessor from llama_index.core.llms import ChatMessage, MessageRole from llama_index.core.postprocessor import ( + SentenceTransformerRerank, SimilarityPostprocessor, ) from llama_index.core.storage import StorageContext @@ -113,16 +114,24 @@ def _chat_engine( context_filter=context_filter, similarity_top_k=self.settings.rag.similarity_top_k, ) + node_postprocessors = [ + MetadataReplacementPostProcessor(target_metadata_key="window"), + SimilarityPostprocessor( + similarity_cutoff=settings.rag.similarity_value + ), + ] + + if settings.rag.rerank.enabled: + rerank_postprocessor = SentenceTransformerRerank( + model=settings.rag.rerank.model, top_n=settings.rag.rerank.top_n + ) + node_postprocessors.append(rerank_postprocessor) + return ContextChatEngine.from_defaults( system_prompt=system_prompt, retriever=vector_index_retriever, llm=self.llm_component.llm, # Takes no effect at the moment - node_postprocessors=[ - MetadataReplacementPostProcessor(target_metadata_key="window"), - SimilarityPostprocessor( - similarity_cutoff=settings.rag.similarity_value - ), - ], + node_postprocessors=node_postprocessors, ) else: return SimpleChatEngine.from_defaults( diff --git a/private_gpt/settings/settings.py b/private_gpt/settings/settings.py index 5896f00d6..bc03e30aa 100644 --- a/private_gpt/settings/settings.py +++ b/private_gpt/settings/settings.py @@ -284,15 +284,31 @@ class UISettings(BaseModel): ) +class RerankSettings(BaseModel): + enabled: bool = Field( + False, + description="This value controls whether a reranker should be included in the RAG pipeline.", + ) + model: str = Field( + "cross-encoder/ms-marco-MiniLM-L-2-v2", + description="Rerank model to use. Limited to SentenceTransformer cross-encoder models.", + ) + top_n: int = Field( + 2, + description="This value controls the number of documents returned by the RAG pipeline.", + ) + + class RagSettings(BaseModel): similarity_top_k: int = Field( 2, - description="This value controls the number of documents returned by the RAG pipeline", + description="This value controls the number of documents returned by the RAG pipeline or considered for reranking if enabled.", ) similarity_value: float = Field( None, description="If set, any documents retrieved from the RAG must meet a certain match score. Acceptable values are between 0 and 1.", ) + rerank: RerankSettings class PostgresSettings(BaseModel): diff --git a/pyproject.toml b/pyproject.toml index d56899987..21d68a85f 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -37,6 +37,11 @@ asyncpg = {version="^0.29.0", optional = true} # Optional Sagemaker dependency boto3 = {version ="^1.34.51", optional = true} + +# Optional Reranker dependencies +torch = {version ="^2.1.2", optional = true} +sentence-transformers = {version ="^2.6.1", optional = true} + # Optional UI gradio = {version ="^4.19.2", optional = true} @@ -57,6 +62,7 @@ vector-stores-qdrant = ["llama-index-vector-stores-qdrant"] vector-stores-chroma = ["llama-index-vector-stores-chroma"] vector-stores-postgres = ["llama-index-vector-stores-postgres"] storage-nodestore-postgres = ["llama-index-storage-docstore-postgres","llama-index-storage-index-store-postgres","psycopg2-binary","asyncpg"] +rerank-sentence-transformers = ["torch", "sentence-transformers"] [tool.poetry.group.dev.dependencies] black = "^22" diff --git a/settings.yaml b/settings.yaml index 87a63ef4f..ce6a2b9fa 100644 --- a/settings.yaml +++ b/settings.yaml @@ -47,6 +47,10 @@ rag: #This value controls how many "top" documents the RAG returns to use in the context. #similarity_value: 0.45 #This value is disabled by default. If you enable this settings, the RAG will only use articles that meet a certain percentage score. + rerank: + enabled: false + model: cross-encoder/ms-marco-MiniLM-L-2-v2 + top_n: 1 llamacpp: prompt_style: "mistral"