Setup

Docker ( 🚰 under construction. tested on Ubuntu LTS)

docker pull su77ungr/casalioy:stable

docker run -it -p 8501:8501 --shm-size=16gb su77ungr/casalioy:stable /bin/bash

for GPU support of stable use casalioy:gpu (unstable)

All set! Proceed with ingesting your dataset

Build it from source

First install all requirements:

python -m pip install poetry
python -m poetry config virtualenvs.in-project true
python -m poetry install
. .venv/bin/activate
python -m pip install --force streamlit sentence_transformers  # Temporary bandaid fix, waiting for streamlit >=1.23
pre-commit install

If you want GPU support for llama-ccp:

pip uninstall -y llama-cpp-python
CMAKE_ARGS="-DLLAMA_CUBLAS=on" FORCE_CMAKE=1 pip install --force llama-cpp-python

Edit the example.env to fit your models and rename it to .env

# Generic
MODEL_N_CTX=1024
TEXT_EMBEDDINGS_MODEL=sentence-transformers/all-MiniLM-L6-v2
TEXT_EMBEDDINGS_MODEL_TYPE=HF  # LlamaCpp or HF
USE_MLOCK=true

# Ingestion
PERSIST_DIRECTORY=db
DOCUMENTS_DIRECTORY=source_documents
INGEST_CHUNK_SIZE=500
INGEST_CHUNK_OVERLAP=50

# Generation
MODEL_TYPE=LlamaCpp # GPT4All or LlamaCpp
MODEL_PATH=eachadea/ggml-vicuna-7b-1.1/ggml-vic7b-q5_1.bin
MODEL_TEMP=0.8
MODEL_STOP=[STOP]
CHAIN_TYPE=stuff
N_RETRIEVE_DOCUMENTS=100 # How many documents to retrieve from the db
N_FORWARD_DOCUMENTS=6 # How many documents to forward to the LLM, chosen among those retrieved

This should look like this

└── repo
      ├── startLLM.py
      ├── casalioy
      │   └── ingest.py, load_env.py, startLLM.py, gui.py, ...
      │   └── misc/ 
      ├── source_documents
      │   └── sample.csv
      │   └── ...
      ├── models
      │   ├── ggml-vic7b-q5_1.bin
      │   └── ...
      └── .env, Dockerfile, ...

👇 Update your installation!

  git pull && poetry install

Ingesting your own dataset

To automatically ingest different data types (.txt, .pdf, .csv, .epub, .html, .docx, .pptx, .eml, .msg)

This repo includes dummy files inside source_documents to run tests with.

python casalioy/ingest.py # optional <path_to_your_data_directory>

Optional: use y flag to purge existing vectorstore and initialize fresh instance

python casalioy/ingest.py # optional <path_to_your_data_directory> y

This spins up a local qdrant namespace inside the db folder containing the local vectorstore. Will take time, depending on the size of your document. You can ingest as many documents as you want by running ingest, and all will be accumulated in the local embeddings database. To remove dataset simply remove db folder.

Ask questions to your documents, locally!

In order to ask a question, run a command like:

python casalioy/startLLM.py

And wait for the script to require your input.

> Enter a query:

Hit enter. You'll need to wait 20-30 seconds (depending on your machine) while the LLM model consumes the prompt and prepares the answer. Once done, it will print the answer and the 4 sources it used as context from your documents; you can then ask another question without re-running the script, just wait for the prompt again.

Note: you could turn off your internet connection, and the script inference would still work. No data gets out of your local environment.

Type exit to finish the script.

Chat inside GUI (new feature)

Introduced by @alxspiker -> see #21

streamlit run casalioy/gui.py

LLM options

Leaderboard

List of available open LLMs HuggingFace

models outside of the GPT-J ecosphere (work out of the box)

🪢 avoid using non-v3 models when using other quantization than q5 (LlamaCpp introduced v3 for ggml)

models inside of the GPT-J ecosphere

all the supported models from here (custom LLMs in Pipeline)

Convert GGML model to GGJT-ready model v1 (for truncation error or not supported models)

1. Download ready-to-use models

Browse Hugging Face for models

2. Convert locally

python casalioy/misc/convert.py see discussion

Pipeline

Selecting the right local models and the power of LangChain you can run the entire pipeline locally, without any data leaving your environment, and with reasonable performance.

• ingest.py uses LangChain tools to parse the document and create embeddings locally using LlamaCppEmbeddings. It then stores the result in a local vector database using Qdrant vector store.

• startLLM.py can handle every LLM that is llamacpp compatible (default GPT4All-J). The context for the answers is extracted from the local vector store using a similarity search to locate the right piece of context from the docs.

Disclaimer

The contents of this repository are provided "as is" and without warranties of any kind, whether express or implied. We do not warrant or represent that the information contained in this repository is accurate, complete, or up-to-date. We expressly disclaim any and all liability for any errors or omissions in the content of this repository.

By using this repository, you are agreeing to comply with and be bound by the above disclaimer. If you do not agree with any part of this disclaimer, please do not use this repository.