A chatbot expert in second-hand item price negotiation based on fuzzy logic.
This project is a proof of concept designed for a class assigment for the IDIF subject (Modelling with uncertainity, Fuzzy Logic and Soft Computing) within the TECI Master degree from the Polytechnic University of Madrid and the Complutense University of Madrid.
Important
Legal Disclaimer: This project is purely academic and educational. Wallabot is not affiliated with, sponsored by, or related to Wallapop S.L. in any way.
Note
Although the documentation and code are in English, the interface and prompts are written in Spanish since this was the language for the lecture. Sorry for the inconvenience!
Chatbot UI
Control Panel: Main Control Panel (left) and Membership Function (right)
-
Create a
.envfile following the structure inexample.envand add your OpenAI API token. If you don't have one you can easily create one here. -
Run the app with
uv:
uv run streamlit run main.py- Access the app in http://localhost:8501 and enjoy the haggling! 🥸$$.
The chatbot's behavior is controlled by a fuzzy inference system.
-
Conversation Duration: The total number of user-bot interactions.
-
User's Tone: (Friendly, Neutral, or Aggressive). I used the bart-large-mnli zero-shot classification model to determine the probability distribution of the three classes. The tone score is then calculated as:
$$Tone = 0 \cdot P(\text{"Friendly"}) + 5 \cdot P(\text{"Neutral"}) + 10 \cdot P(\text{"Aggressive"})$$ 0 indicates a friendly tone, while values closer to 10 represent higher aggression.
-
Price difference: User's price offer is extracted from the user's message using regular expressions and then compared to the original price to calculate the relative difference (%).
The output is the Degree of Acceptance. Depending on this degree, a specific strategy is selected and injected into the system prompt:
- Very High: Accept offer.
- High: Make a counteroffer.
- Low: Maintain the current price.
- Very Low: Reject the offer and end the negotiation.
The rules were defined following well-known negotiation principles (inspired by works like Dale Carnegie's How to Win Friends and Influence People), such as "Never reward aggression with concessions" and "Prioritize good faith and mutual benefit". The following table displays the 17 rules that define the chatbot's decision making:
| Buyer's Tone | Relative Price Difference | Negotiation Duration | Action |
|---|---|---|---|
| Friendly | Low | Short | Accept |
| Friendly | Low | Medium | Accept |
| Friendly | Low | Long | Counteroffer |
| Friendly | Medium | Short | Counteroffer |
| Friendly | Medium | Medium | Counteroffer |
| Friendly | Medium | Long | Maintain |
| Friendly | High | Short | Counteroffer |
| Friendly | High | Medium | Maintain |
| Friendly | High | Long | Reject |
| Neutral | Low | Short | Accept |
| Neutral | Low | Medium | Accept |
| Neutral | Low | Long | Counteroffer |
| Neutral | Medium | Short | Counteroffer |
| Neutral | Medium | Medium | Maintain |
| Neutral | Medium | Long | Maintain |
| Neutral | High | Short | Maintain |
| Neutral | High | Medium | Maintain |
| Neutral | High | Long | Maintain |
| Aggressive | Low | Short | Maintain |
| Aggressive | Low | Medium | Maintain |
| Aggressive | Low | Long | Reject |
| Aggressive | Medium | Short | Maintain |
| Aggressive | Medium | Medium | Maintain |
| Aggressive | Medium | Long | Reject |
| Aggressive | High | Short | Reject |
| Aggressive | High | Medium | Reject |
| Aggressive | High | Long | Reject |