A plugin integrating a chat interface in ckan with a pydanticai agent that can use all available ckan actions and url_patterns. All actions are done with user aware context. The chat interface uses marked and highightjs to display responses. Chat histories are saved in the local storage of the user. The agent is chat history aware. LLMs to use are configured in the bot/agent.py a section Model & Agent Setup. An Azure Openai and a local impementation using openai compartible api of ollama is implemeneted.

It has a rag_search tool that can facilitate a Milvus vector store if it is set up. Currently it relies on the Azure OpenAI embeddings api and will not work with local deployments. It uses the embedding model text-embedding-3-small to form the search vector. To make use of the documents it returns, the metadata of te vectores should include the dataset and resource ids at least. See the class VectorMeta for expected fields.

Openai Models starting from gpt-35 on work very well. Local LLMs tested with ollama server are listed below.

in general reasoning models dont perform well

A completion endpoint of the LLM model to use with the agent is needed. Currently uses Azure Cognitive Service Integration. can be changed by replacing the client in /bot/agent.py

Compatibility with core CKAN versions:

Suggested values:

• "yes"
• "not tested" - I can't think of a reason why it wouldn't work
• "not yet" - there is an intention to get it working
• "no"

To install the extension:

1. Activate your CKAN virtual environment, for example:

. /usr/lib/ckan/default/bin/activate

2. Use pip to install package

pip install ckanext-chat

3. Add csvtocsvw to the ckan.plugins setting in your CKAN    config file (by default the config file is located at    /etc/ckan/default/ckan.ini).

4. Restart CKAN. For example, if you've deployed CKAN with Apache on Ubuntu:

sudo service apache2 reload

In your env variables set:

CKANINI__CKANEXT__CHAT__COMPLETION_URL="https://your-subscription.openai.azure.com/"
CKANINI__CKANEXT__CHAT__DEPLOYMENT="gpt-4o"
CKANINI__CKANEXT__CHAT__API_TOKEN="your-api-token"

or ckan.ini parameters.

ckanext.chat.completion_url="https://your-subscription.openai.azure.com/"
ckanext.chat.deployment="gpt-4o"
ckanext.chat.api_token="your-api-token"

To not run into api call timeouts the proxy infromt of ckan must be set to allow long running api calls for nginx

proxy_connect_timeout 3600s;
proxy_read_timeout 3600s;
proxy_send_timeout 3000s;
send_timeout 3000;

for production if ure using the official docker containers of ckan the harakiri options must be set. For this edit the start_ckan.sh script:

UWSGI_OPTS="--socket /tmp/uwsgi.sock \
            --wsgi-file /srv/app/wsgi.py \
            --module wsgi:application \
            --http 0.0.0.0:5000 \
            --master --enable-threads \
            --lazy-apps \
            -p 2 -L -b 32768 --vacuum \
            --harakiri-verbose \
            --socket-timeout $UWSGI_HARAKIRI \
            --harakiri $UWSGI_HARAKIRI \
            --http-timeout $UWSGI_HARAKIRI"

set in.env

UWSGI_HARAKIRI="3000"

if your also setup an Milvus vector database for rag search of documents or alike there is options you can set

ckanext.chat.embedding_mode=<embedding model name to request from the embedding api>
ckanext.chat.embedding_api=<api endpoint to send text to and to return an embeding>
ckanext.chat.milvus_url=<url to milvus server>
ckanext.chat.collection_name=<name of milvus collection

. You might need to lookup and change the exact embedding api generation because no api standard applies! If you dont set this options the literature_search agent will rely on the package_search action!

To install ckanext-csvtocsvw for development, activate your CKAN virtualenv and do:

git clone https://github.com/Mat-O-Lab/ckanext-chat.git
cd ckanext-chat
python setup.py develop
pip install -r dev-requirements.txt

To run the tests, do:

pytest --ckan-ini=test.ini

AGPL