LLM Search

The purpose of this package is to offer a convenient question-answering system with a simple YAML-based configuration that enables interaction with multiple collections of local documents. Special attention is given to improvements in various components of the system in addition to LLMs - better document parsing, hybrid search, HyDE enabled search, deep linking, re-ranking, the ability to customize embeddings, and more. The package is designed to work with custom Large Language Models (LLMs) – whether from OpenAI or installed locally.

Features

• Supported formats

  • .md - Divides files based on logical components such as headings, subheadings, and code blocks. Supports additional features like cleaning image links, adding custom metadata, and more.
  • .pdf - MuPDF-based parser.
  • .html, .epub - supported through the Unstructured pre-processor - https://unstructured-io.github.io/unstructured/
  • .docx - custom parser, supports nested tables.

• Supports multiple collection of documents, and filtering the results by a collection.

• Generates dense embeddings from a folder of documents and stores them in a vector database (ChromaDB).

  • The following embedding models are supported:
    • Huggingface embeddings.
    • Sentence-transformers-based models, e.g., multilingual-e5-base.
    • Instructor-based models, e.g., instructor-large.

• Generates sparse embeddings using SPLADE (https://github.com/naver/splade) to enable hybrid search (sparse + dense).

• Supports the "Retrieve and Re-rank" strategy for semantic search, see - https://www.sbert.net/examples/applications/retrieve_rerank/README.html.

  • Besides the originally ms-marco-MiniLM cross-encoder, more modern bge-reranker is supported.

• Supports HyDE (Hypothetical Document Embeddings) - https://arxiv.org/pdf/2212.10496.pdf

  • WARNING: Enabling HyDE (via config OR webapp) can significantly alter the quality of the results. Please make sure to read the paper before enabling.
  • From my own experiments, enabling HyDE significantly boosts quality of the output on a topics where user can't formulate the quesiton using domain specific language of the topic - e.g. when learning new topics.

• Support for multi-querying, inspired by RAG Fusion - https://towardsdatascience.com/forget-rag-the-future-is-rag-fusion-1147298d8ad1

  • When multi-querying is turned on (either config or webapp), the original query will be replaced by 3 variants of the same query, allowing to bridge the gap in the terminology and "offer different angles or perspectives" according to the article.

• Allows interaction with embedded documents, supporting the following models and methods (including locally hosted):

  • OpenAI models (ChatGPT 3.5/4 and Azure OpenAI).
  • HuggingFace models.
  • Llama cpp supported models - for full list see https://github.com/ggerganov/llama.cpp#description
  • AutoGPTQ models (temporarily disabled due to broken dependencies).

• Other features

  • Simple CLI and web interfaces.
  • Deep linking into document sections - jump to an individual PDF page or a header in a markdown file.
  • Ability to save responses to an offline database for future analysis.
  • Experimental API

Demo

Prerequisites

• Tested on Ubuntu 22.04.
• Nvidia GPU is required for embeddings generation and usage of locally hosted models.
• Python 3.10, including dev packages (python3-dev on Ubuntu)
• Nvidia CUDA Toolkit (tested with v11.7) - https://developer.nvidia.com/cuda-toolkit
• To interact with OpenAI models, create .env in the root directory of the repository, containing OpenAI API key. A template for the .env file is provided in .env_template
• For parsing .epub documents, Pandoc is required - https://pandoc.org/installing.html

Virtualenv based installation

git clone https://github.com/snexus/llm-search.git
cd llm-search

# Create a new environment
python3 -m venv .venv 

# Activate new environment
source .venv/bin/activate

# Set variables for llama-cpp to compile with CUDA.

# Assuming Nvidia CUDA Toolkit is installed and pointing to `usr/local/cuda` on Ubuntu

source ./setvars.sh 


# Install the package
pip install . # or `pip install -e .` for development

Quickstart

1) Create a configuration file

To create a configuration file in YAML format, you can refer to the example template provided in sample_templates/generic/config_template.yaml.

The sample configuration file specifies how to load one of the supported locally hosted models, downloaded from Huggingface - https://huggingface.co/TheBloke/airoboros-l2-13B-gpt4-1.4.1-GGUF/resolve/main/airoboros-l2-13b-gpt4-1.4.1.Q4_K_M.gguf

As an alternative uncomment the llm section for OpenAI model.

Sample configuration template

2) Create document embeddings

To create embeddings from documents, follow these steps:

1. Open the command line interface.
2. Run the following command:

llmsearch index create -c /path/to/config.yaml

Based on the example configuration above, executing this command will scan a folder containing markdown and pdf files (/path/to/docments) excluding the files in subfolder1 and subfolder2 and generate a dense embeddings database in the /path/to/embedding/folder directory. Additionally, a local cache folder (/path/to/cache/folder) will be utilized to store embedding models, LLM models, and tokenizers.

The default vector database for dense is ChromaDB, and default embedding model is e5-large-v2 (unless specified otherwise using embedding_model section such as above), which is known for its high performance. You can find more information about this and other embedding models at https://huggingface.co/spaces/mteb/leaderboard.

In addition to dense embeddings, sparse embedding will be generated in /path/to/embedding/folder/splade using SPLADE algorithm. Both dense and sparse embeddings will be used for context search.

3) Update document embeddings

When new files are added or existing documents are changed, follow these steps to update the embeddings:

llmsearch index update -c /path/to/config.yaml

Executing this command will detect changed or new files (based on MD5 hash) and will incrementally update only the changes, without the need to rescan the documents from scratch.

4) Interact with the documents

To interact with the documents using one of the supported LLMs, follow these steps:

1. Open the command line interface.
2. Run one of the following commands:

• Web interface:

Scans the configs and allows to switch between them.

llmsearch interact webapp -c /path/to/config_folder

• CLI interface:

llmsearch interact llm -c /path/to/config.yaml

Based on the example configuration provided in the sample configuration file, the following actions will take place:

• The system will load a quantized GGUF model using the LlamaCpp framework. The model file is located at /storage/llm/cache/airoboros-l2-13b-gpt4-1.4.1.Q4_K_M.gguf
• The model will be partially loaded into the GPU (30 layers) and partially into the CPU (remaining layers). The n_gpu_layers parameter can be adjusted according to the hardware limitations.
• Additional LlamaCpp specific parameters specified in model_kwargs from the llm->params section will be passed to the model.
• The system will query the embeddings database using hybrid search algorithm using sparse and dense embeddings. It will provide the most relevant context from different documents, up to a maximum context size of 4096 characters (max_char_size in semantic_search).
• When displaying paths to relevant documents, the system will replace the part of the path /storage/llm/docs/ with obsidian://open?vault=knowledge-base&file=. This replacement is based on the settings substring_search and substring_replace in semantic_search->replace_output_path.

API (experimental)

To launch an api, supply a path config file in the FASTAPI_LLM_CONFIG environment variable and launch llmsearchapi

FASTAPI_LLM_CONFIG="/path/to/config.yaml" llmsearchapi