Converse with an LLM provider of your choice. Store your documents and chat history in a Solid pod.
This project builds on the work from ChatDocs-Streamlit.
# local installation of the package:
pip install .
# editable install:
pip install -e .
# editable install with optional dependencies:
pip install -e .[llm]If you want to run the LLM service with GPU acceleration, then you'll also need to:
- Install PyTorch with CUDA enabled by following the instructions here
pip install ctransformers[cuda]
If running locally, you don't need to do anything. If deploying remotely, you may need to define
the WEBSITE_HOSTNAME environment variable to point to your chat app url (Azure hosting does
this automatically, on a Google Cloud VM you need to export the variable, for example: export WEBSITE_HOSTNAME=34.105.136.95:8501). Running over https is required when not on localhost.
If running locally or on a VM, navigate into the llm_service directory and copy the provided .env.sample file to a .env file. Then fill in the values with desired details for a retrieval service provider.
If deploying remotely, instead of using the .env file, configure your cloud provider setup to include the environment variables from .env.sample.
Example .env
RETRIEVAL_SERVICE_IDP=https://example.org
RETRIEVAL_SERVICE_NAME=socialgenpod_retriever
RETRIEVAL_SERVICE_EMAIL=retriever@example.org
RETRIEVAL_SERVICE_PASSWORD=supersecret
RETRIEVAL_SERVICE_WEBID=https://example.org/socialgenpod_retriever⚠ Caveat: due to non-standardisation of client credentials authentication, the retrieval service provider must be registered with a Community Solid Server (Node Solid Servers like https://solidcommunity.net and Enterprise Solid Servers are currently not supported). This only applies to the retrieval service provider, the chat app users can use any Solid server as their identity provider.
For Solid Servers with Web Access Control (WAC / ACL), you can use either Mashlib or Penny. Access Control Policies (ACP) do not currently have a GUI for editing permissions.
Mashlib: https://solidos.github.io/mashlib/dist/browse.html (see https://github.com/SolidOS/userguide#manage-your-trusted-applications for guidance).
Penny: https://penny.vincenttunru.com/. Point Penny to browse the pod in which the files you would like to share are stored. Then login as the pod owner. Navigate to the resource you would like to share and, under the "Linked Resources" section, press "Add Access Control List". This will create a new .acl file with all permissions granted to the pod owner and nothing else. Create a "new thing" in the ACL file, choose "Convert to Access Control" and grant only the "Read" permission to the webid of the retrieval service provider, e.g. https://example.org/socialgenpod_retriever/profile/card#me.
If you would like to add different language models or change the configuration, create a file called genpod.yml in your working directory before running genpod-llm. This will be merged with the configuration in llm_service/data/genpod.yml -- see that file for example syntax.
For example, add download: True in your configuration to enable downloading models specified in the configuration file -- otherwise only already locally stored models will be possible to use.
You can also change the hostname and port the service runs on.
For other configuration, such as adding GPU acceleration, see https://github.com/Vidminas/chatdocs-streamlit. The configuration file works the same way.
Example genpod.yml with GPU acceleration
host: 10.154.0.4
embeddings:
model_kwargs:
device: cuda
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
gpu_layers: 50
- model_framework: huggingface
model: RWKV/v5-EagleX-v2-7B-HF
model_kwargs:
trust_remote_code: true
device: 0
- model_framework: huggingface
model: NousResearch/Hermes-2-Pro-Mistral-7B
pipeline_kwargs:
max_new_tokens: 200
device: 0You can deploy the providers on a cloud hosting service or run them locally. When running locally, you could use network tunneling as a quick way to make them available for other people (e.g., using ngrok, serveo, or localtunnel).
Open chat interface by running genpod-chat.
Run the LLM service provider using genpod-llm.
On opening the chat app, log in with Solid. You will be prompted to configure the chatbot service by providing endpoints of the Retrieval Service Provider and LLM Provider. These may or may not be hosted at the same URL. Optionally, you can add in the path to resources that should be included in retrieval. The app will check whether it can establish a connection to both providers and, if a document location is provided, whether the Retrieval Service Provider is able to access it.
See our demo paper (available on arXiv: SocialGenPod: Privacy-Friendly Generative AI Social Web Applications with Decentralised Personal Data Stores) for the motivation and overview.
This is an example, simplified flow of network requests when a user, Bob, uses the chat app (assuming the service providers and access permissions have already been configured). It details the example use case described in the demo paper.
---
title: SocialGenPod initialisation
config:
mirrorActors: false
---
sequenceDiagram
actor Bob
participant App as ChatApp
participant BIDP as Bob's Identity Provider
participant BPod as Bob's Solid Pod
participant ConfigPod as Any Solid Pod
participant Retriever as Retrieval Service
participant LLM as LLM Service
participant APod as Alice's Solid Pod
Bob ->>+ App: login()
App ->+ BIDP: oidc_login()
BIDP ->> Bob: login screen
Bob ->> BIDP: login(username, password)
BIDP -->>- App: oidc_token
%% App ->>+ BPod: getOrCreateSocialGenPodConfig(oidc_token)
%% BPod ->>- App: config
App ->> Bob: show_config()
alt
Bob ->> App: config_url
App ->>+ ConfigPod: get_socialgenpod_config(oidc_token)
ConfigPod -->>- App: retrieval_service, llm_provider, documents_location
else
Bob ->> App: retrieval_service, llm_provider, documents_location
end
opt documents_location is not None
App ->>+ Retriever: test_access(retrieval_service, oidc_token, documents_location)
critical Prepare vector retrieval DB
Retriever ->> Retriever: uris = discover_document_uris()
loop for uri in uris
Retriever ->>+ APod: download_resource(uri)
APod -->>- Retriever: content
Retriever ->> Retriever: add(content)
end
option error
Retriever ->> Retriever: log(error)
end
Retriever -->>- App: ok
end
App ->>+ LLM: test_access(llm_provider, oidc_token)
LLM -->>- App: ok
App ->>+ BPod: getChatMessageHistory(oidc_token)
BPod -->>- App: chat_history
---
title: SocialGenPod communication
config:
mirrorActors: false
---
sequenceDiagram
actor Bob
participant App as ChatApp
participant BPod as Bob's Solid Pod
participant Retriever as Retrieval Service
participant LLM as LLM Service
Bob ->>+ App: chat_input(prompt)
opt chat_history length > 1
App ->>+ LLM: summarise(prompt, chat_history)
LLM -->>- App: prompt = condensed_prompt
end
App ->>+ BPod: store_message(prompt)
deactivate BPod
opt documents_location is not None
App ->>+ Retriever: find_relevant_context(selected_retriever, documents_location, prompt)
Retriever -->>- App: relevant_documents
end
App ->>+ LLM: chat_completion(selected_llm, prompt, relevant_documents)
LLM -->>- App: response
App ->>+ BPod: store_message(response)
deactivate BPod
App -->>- Bob: response
chat_app/data/turtle.png is from https://emojipedia.org/mozilla/firefox-os-2.5/turtle.
If this work is relevant to your research or if you use it in your work, please consider citing
@inproceedings{socialgenpod2024,
title = {{{SocialGenPod}}: {{Privacy-Friendly Generative AI Social Web Applications}} with {{Decentralised Personal Data Stores}}},
booktitle = {Companion {{Proceedings}} of the {{ACM Web Conference}} 2024},
author = {{Vidminas Vizgirda} and {Rui Zhao} and {Naman Goel}},
publisher = {{ACM}},
address = {{Singapore}},
doi = {10.1145/3589335.3651251},
isbn = {979-8-4007-0172-6/24/05}
}