A RAG-enabled workbench to index files and chat with an LLM about their contents.

Download and install Ollama, a framework to interact with LLMs.

After installation, run Ollama:

ollama serve

In order to configure the model and the application, in a new terminal, run:

make setup

make

A page to load context files and interact with the LLM will open in your browser.

• One-shot prompts to LLM.
• File indexing for context querying.
• Prompt tools to assist with prompt construction and context gathering.
• Replaying of a set of prompts, either from the current prompt history or a text file.
• Displaying of all prompts and responses in the chat container.
• Download of all or only the last chat messages in a chat.txt text file.

Tools are used directly in the chat message input box.

:<label> /<tool> <prompt text, can contain {response:*} for replacement>

Given a previous JSON response, it's possible to use the /template tool to create a template that will be processed using the JSON data.

Using the following JSON as the last response in the prompt history:

{
    "name": "User"
}

The prompt below will generate a response using the JSON as input data in the context variable:

/template Name: {{context.name}}

The response to the prompt will be:

Name: User

{% set variable_name = context %}

{{context.field_date|parse_date|format_date("%d/%m/%y %H:%M:%S")}}

{% if context.field_boolean %}
Value if True
{% else %}
Value if False
{% endif %})

 {% for item in context.list %}
 {{item.field}}
 {% endfor %}

Please refer to Jinja and jinja2_iso8601 documentations for more details on templating.

In case you want to use API mocking to test context retrieval from endpoints, it's possible to use the JSON Server package for mocking.

Having Node.js/NPM installed, run the the following command to install dependencies:

make setup/server

Add the JSON you want to use as mock data in the db.json file and run the server in a new terminal with the command below:

make run/server

The server will be accessible in http://localhost:3000/, with the root nodes of the JSON file as URL paths (e.g. in the demo db.json file) there's a data root node, which can be accessible through http://localhost:3000/data.

By default, the workbench uses the LLM model Llama3.1.

To change the LLM model used by the workbench, update the FROM parameter in contextualized_assistant.model file by a model available in the Ollama library.

1. The buttons in the screen are not always disabled during operations. Please be aware that clicking on different buttons during actions may lead to unintended consequences.
2. The download of chat history may not work during first attempt.
3. Complex Excel/.xlsx files may not be loadable due to format incompatibility with openpyxl.
4. During replay, the scrolling may not be automatic.