README.md

---
title: Audit Assistant
emoji: 🧐
colorFrom: yellow
colorTo: gray
sdk: gradio
sdk_version: 4.39.0
app_file: app.py
fullWidth: true
pinned: false
startup_duration_timeout: 1h
models:
- BAAI/bge-m3
- BAAI/bge-reranker-v2-m3
- meta-llama/Llama-3.1-8B-Instruct

short_description: Ask any questions to the Audit reports
license: apache-2.0
---
### Technical Documentation of the system in accordance with EU AI Act.
System Name: Audit Assistant

Provider / Supplier: GIZ Data Service Center

As of: July 2025


1. General Description of the System
Audit Assistant is an AI-powered tool for exploring and understanding Uganda's audit reports. This tool leverages advanced language models to help you get clear and structured answers based on audit publications. It combines a generative language assistant with a knowledge base implemented via Retrieval-Augmented Generation (RAG).


2. Model's Used 

    Generative LLM
    -   Model Name: [meta-llama/Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct)
    -	Model Source API: [Nebius AI](https://studio.nebius.com/)

    Retriever/Embedding
    -   Model Name: [BAAI/bge-m3](https://huggingface.co/BAAI/bge-m3)
    -   Model Source: Local Instance 

    Re-ranker:
    -   Model Name: [BAAI/bge-reranker-v2-m3](https://huggingface.co/BAAI/bge-reranker-v2-m3)
    -   Model Source: Local Instance

3. Model Training Data:
All the models mentioned above are being consumed without any fine-tuning or training being performed by the developer team of Audit Assistant tool. And hence there is no training data which had been used by the development team of Audit Assistant tool.


4. Knowledge Base (Retrieval Component)

    -   Data Sources: Public, non-personal audit reports relased by [**Office of the Auditor General (OAG) of Uganda**](https://www.oag.go.ug/welcome)
    -   Embedding Model: [BAAI/bge-m3](https://huggingface.co/BAAI/bge-m3)
    -   Embedding Dimension: 1024
    -   Vector Database: Qdrant (via API)
    -   Framework: Langchain (custom RAG pipeline)
    -   Top-k: 5 relevant text segments per query

5. System Limitations and Non-Purposes
    -	The system does not make autonomous decisions.
    -	No processing of personal data except for the usage statistics as mentioned in Disclaimer.
    -	Results are intended for orientation only – not for legal or political advice.

6. Transparency Towards Users
    -	The user interface clearly indicates the use of a generative AI model.
    -	An explanation of the RAG method is included.
    -   We collect usage statistics as detailed in Disclaimer tab of the app along with the explicit display in the user interface of the tool.
    -	Feedback mechanism available (via https://huggingface.co/spaces/GIZ/audit_assistant/discussions/new).

7. Monitoring, Feedback, and Incident Reporting
    -	User can provide feedback via UI by giving (Thumbs-up or down to AI-Generated answer). Alternatively for more detailed feedback please use https://huggingface.co/spaces/GIZ/audit_assistant/discussions/new to report any issue.
    -	Technical development is carried out by the GIZ Data Service Center.
    -	No automated bias detection – but low risk due to content restrictions.


8. Contact: For any questions, please contact via https://huggingface.co/spaces/GIZ/audit_assistant/discussions/new or send us email to dataservicecenter@giz.de