Weconomics AI assistant
AI & Data
Semester programme:Open Learning/Innovation
Client company:Weconomics
Project group members:Kamen Trifonov
Marek Mitala
Sahar Kazemi
Radja Dalimonthee
Project description
Our stakeholder presented us with a challenge related to company documentation. Namely, the amount of documents makes it difficult to find information quickly and accurately when needed. As such, our task was to make a solution that allows for quick lookup and match of data -> relevant document and present it as such.
Context
The stakeholder wanted to use this project as a proof of concept for a product he could sell to companies. It evolved into an AI assistant that could accomplish that task and any additional capabilities we added to the core was welcome, so long as it expanded the original vision.
Results
Through weekly alignment meetings and continued development, our project came to be an AI assistant platform based on GLM 4.5+ utilizing open source models and technologies to fil in the stopgaps and expand our core feature set. We utilize Marker for compatible document conversion from a bucket. The resulting HTML file is uploaded to Pinecone for automatic vector conversion and storage. The core AI model and function set is hosted in Supabase. Retrieval of documents happens through an MCP server, client and corresponding requirements like wrappers built based on the Pinecone SDK. We built a websearch microservice based on the duckduckgo API using an MCP server. We built-in a fallback to Grok as a secondary model utilizing llama.