Model Card: Kweli

Model Details

Model Name: Kweli

Version: 0.1(Proof of Concept)

Model Type: Large Language Model (LLM)

Developed By:ALX Africa, Data Science Department

Model Description: Kweli is an agentic support agent built on the Supervisor agent architecture. It has a dual purpose: to answer questions about ALX and its impact, and to query the impact database for information.

Intended Users : Kweli is designed to be used by Fred, and a couple of executives at ALX. Kweli has access to the 2024 Mastercard proposal document, 2025 ALX pitch deck, and ALX onboarding course launching soon.

Last Updated: 05/03/2025

Intended Use

Primary Use Cases:

• Query the impact database for up to date information about ALX and its impact.
• Answer questions about ALX past, future, plans, mission, vision, values, and more.

Potential Applications:

• The SQL agent has a great capability to improve the data-business process. In future iterations, we hope that Kweli can be asked any business question related to data, and Kweli will search all of our data sources to bring answers to descision makers in minutes.

Data and architecture

Models: The model is based on OpenAi's GPT-4o model and is augmented by ALX documents and policies.

Architecture: Kweli is built on the Langgraph platform, and has two agents: a SQL agent and a RAG agent. The SQL agent is designed to answer questions about our impact data, and the RAG agent is designed to collect information from our document store.

Performance expectation

Proof of concept stage:

• Answer basic queries about impact numbers: Assessed on a test set of course-related and support queries.
• Answer queries about ALX past, future, plans, mission, vision, values, and more.
• Optimised for quality at the cost of speed: Kweli can take 10-20s to respond. It has agents discussing the best way to answer the question, and then the SQL agent or RAG agent fetches the information the agent needs, and it crafts a response.

Limitations

Unreliability: Kweli is in a PoC phase. Ideally we use it, get feedback and iterate. It will likely not be accurate at the start.

Incomplete Information: THe underlying database is still changing, so the agent may not keep up with the changes.

Slow response time: Kweli can take 10-20s to respond. It has agents discussing the best way to answer the question, and then the SQL agent or RAG agent fetches the information the agent needs, and it crafts a response.

Privacy

Conversations are stored on a secure LangSmith platform to improve our product. Currently, Stefan Coetzee and Ahmed Abdulakhir have access to these conversations, and keep all conversations and data private.

The data handing is compliant with data protection regulations, with data protection agreements signed by all providers.

Usage Guidelines

Target Audience: In the PoC phase, the agent is designed to be used by Fred, and a couple of executives. Kweli has access to the 2024 Mastercard proposal, 2025 ALX pitch deck, and ALX onboarding course launching soon.