This repository contains an Agentic RAG (Retrieval Augmented Generation) System developed for the City of Amsterdam and used to explore the use of this technology for improving the handling of citizen reports. The solution automates responses to provide fast, empathetic, and context-specific feedback, reducing delays and dissatisfaction. By automating these cases, the workload for municipal staff is reduced, allowing more focus on complex issues.
Example interaction with the system.
When a melding is received, the system retrieves relevant information
from various sources, with the aim of providing useful information
about the issue, whether, when and how it would be resolved
and what the citizen can expect.
Disclaimer: Due to the sensitive nature of some of the underlying data sources (e.g. the historical citizen reports), running the system requires incorporating own data and making the necessary adjustments. We hope, however, that its architecture, code and underlying prompts can serve as inspiration for similar use cases.
Amsterdam's Meldingen system is an online system which allows citizens to report issues in public spaces. Citizens can report issues such as rubbish or a maintenance issue on the street or in a park, a dangerous traffic situation or disturbance from people or cafes. According to data from 2023, the City receives ± 400.000 citizen reports per year.
With 25-30% of reports requiring no (new) action from the municipality, it is expected that automatically collecting information from internal systems and providing the citizen with details about the issue would have a significant impact on users. The solution would provide citizens with fast and accurate responses, it will make the process of handling reports more transparent, it will create user-friendly interaction, and - in the long term - ensure better service for reports that do require human handling.
More information about the envisioned solution and its impact on users can be found in the full report (English|Dutch).
Our solution covers the most relevant components of an agentic RAG architecture. It focuses primarily on the implementation of diverse tools which can help us assess the feasibility of different scenarios and the adaptability and scalability of the system.
Overview of our Agentic RAG architecture
In the core of the system is the central agent. It serves as an orchestrator, which receives instructions regarding the task and the report and collects all relevant details about a report, such as the type of problem and the location. It then independently determines which tool is suitable for gathering information.
In its configuration, the agent receives instructions about its task, different guidelines related to the resolution of reports, as well as communication rules. Furthermore, the central agent is provided with an explanation of the available tools and their functionalities to use them effectively for handling reports. Finally, the central agent is instructed to follow a structured reasoning process, documenting intermediate steps such as the report, selected tools, gathered information, and preliminary conclusions, ensuring a comprehensive overview of actionable information.
The reasoning processes and effectively combining all gathered data is essentially performed by an LLM. In most experiments we use own deployment of GPT-4o, however, it is possible to optimize the system for the use of any open-source model. This does require adjusting the templates (especially the scratchpad) and configuration (including some of the parameters in the langchain agent initialization) to correspond to the desired model.
More information about the agent, as well as the individual tools can be found in the full report (English|Dutch).
- Clone this repository:
git clone https://github.com/Amsterdam-AI-Team/citizen-signals-rag-chatbot-system.git- Install pyaudio driver:
sudo apt-get install portaudio19-dev- Install all dependencies:
pip install -r requirements.txtThe code has been tested with Python 3.10.0 on Linux.
Disclaimer: This repository is intended primarily for transparency and inspiration. Due to the sensitive nature of some of the underlying data sources (e.g. the historical citizen reports), running the system requires incorporating own data and making the necessary adjustments. This could hinder the usage of the system outside of the secure municipal environment.
Before use, please adjust all paths in the config.py file.
You can also decide whether to track_emissions,
select a different embedding_model_name (used for the retrieval of historical reports),
decide on a model in the model_dict.
You can also adjust all agent templates according to own preferences.
Furthermore, you will need to add a my_secrets.py with an API_KEYS object containing
openai_azurekey: an OpenAI API key corresponding to the endpoint in the config.pyco2-signalkey (iftrack_emissions=True)
You can use the chat interface by running the app located in the source directory using the following command:
python src/app.pyThis will start the chatbot on your localhost, allowing you to interact with it via a web interface. By default you can access the app locally on port 5000 (127.0.0.1:5000).
Alternatively, you can directly call the central agent by providing the report, address and any additional information. To do that, adjust the examples on the bottom of the central_agentic_agent.py script use the following command:
python src/central_agentic_agent.pyIf you wish to run this code via Azure ML services, you can open the repository and run app.py in VS Desktop mode. This will ensure the localhost application works properly.
For internal use of the original meldingen data, mount the corresponding storage container using:
sh blobfuse_meldingen.shWe welcome contributions! Feel free to open an issue, submit a pull request, or contact us directly.
This repository was created by Amsterdam Intelligence for the City of Amsterdam.
This project is licensed under the terms of the European Union Public License 1.2 (EUPL-1.2).