A fork of ChatDocs
Chat with your documents offline using AI. No data leaves your system. Internet connection is only required to install the tool and download the AI models. It is based on PrivateGPT but has more features.
Contents
- Supports GGML/GGUF models via CTransformers
- Supports 🤗 Transformers models
- Web UI
- GPU support
- Highly configurable via
chatdocs.yml
Show supported document types
Extension | Format |
---|---|
.csv |
CSV |
.docx , .doc |
Word Document |
.enex |
EverNote |
.eml |
|
.epub |
EPub |
.html |
HTML |
.md |
Markdown |
.msg |
Outlook Message |
.odt |
Open Document Text |
.pdf |
Portable Document Format (PDF) |
.pptx , .ppt |
PowerPoint Document |
.txt |
Text file (UTF-8) |
Run pip install git+https://github.com/Vidminas/chatdocs-streamlit.git
- Install PyTorch with CUDA enabled by following the instructions here.
pip install ctransformers[cuda]
pip install git+https://github.com/Vidminas/chatdocs-streamlit.git
If pip takes too long to resolve dependency versions, you can also use pip install git+https://github.com/Vidminas/chatdocs-streamlit.git --use-deprecated=legacy-resolver
. This may result in some dependency version conflicts, but should be fine to ignore (some libraries just haven't updated the supported version bounds for their dependencies).
Download the AI models using:
chatdocs download
Now it can be run offline without internet connection.
Add a directory containing documents to chat with using:
chatdocs add /path/to/documents
The processed documents will be stored in
db
directory by default.
Chat with your documents using:
chatdocs ui
Open http://localhost:8501 in your browser to access the web UI.
It also has a nice command-line interface:
chatdocs chat
All the configuration options can be changed using the chatdocs.yml
config file. Create a chatdocs.yml
file in some directory and run all commands from that directory. For reference, see the default chatdocs.yml
file.
You don't have to copy the entire file, just add the config options you want to change as it will be merged with the default config. For example, see tests/fixtures/chatdocs.yml
which changes only some of the config options.
To change the embeddings model, add and change the following in your chatdocs.yml
:
embeddings:
model: hkunlp/instructor-large
Note: When you change the embeddings model, delete the
db
directory and add documents again.
You can configure multiple LLMs to use for chatdocs. The command line interface uses the first one from the list. The UI provides radio buttons to select which one to use.
Each model in the list must specify which framework to use: either CTransformers (GGML/GGUF) or 🤗 Transformers.
To add more models, use the following template in your chatdocs.yml
:
llms:
- model_framework: ctransformers
model: TheBloke/orca_mini_3B-GGML
model_file: orca-mini-3b.ggmlv3.q4_0.bin
model_type: llama
config:
context_length: 1024
max_new_tokens: 256
- model_framework: huggingface
model: TheBloke/Wizard-Vicuna-7B-Uncensored-HF
pipeline_kwargs:
max_new_tokens: 256
CTransformers requires specifying model_type
(between llama, gpt2, gpt3, falcon, ...).
Note: When you add a new model for the first time, run
chatdocs download
to download the model before using it.
You can also use an existing local model file, for example:
llms:
- model_framework: ctransformers
model: /path/to/ggml-model.bin
model_type: llama
Finally, if you wish to compare results with an OpenAI model, you can add:
- model_framework: openai
model: gpt-3.5-turbo-0613
openai_api_key: YOUR-KEY-HERE (starting with sk-)
This is for testing purposes -- if you use the OpenAI models, your documents and chat data will be sent over the API, unlike with local LLMs.
To enable GPU (CUDA) support for the embeddings model, add the following to your chatdocs.yml
:
embeddings:
model_kwargs:
device: cuda
To enable GPU (CUDA) support for a CTransformers (GGML/GGUF) model, add the following to your chatdocs.yml
:
llms:
- model_framework: ctransformers
# ...
config:
gpu_layers: 50
To enable GPU (CUDA) support for the 🤗 Transformers model, add the following to your chatdocs.yml
:
llms:
- model_framework: huggingface
# ...
device: 0