Code for our MSc thesis, a continuation of our paper "LLMs as Proxy Survey Participants With RAG", by Elias Torjani, Airidas Brikas, and Daniel Hardt (our BSc thesis)

Check out our [abstract-length] paper on it: Market research via persona-induced Large Language Models, or see our poster below as a TL;DR

---

How to reproduce our experiments with your own data

1. Export your chat messages from Facebook, Instagram, and/or WhatsApp (instructions below)
2. Take the surveys to constitute target responses, for the LLMs proxying you in the same surveys.
3. Clone this repository
4. Download Ollama, llama.cpp+gguf-models vLLM [WIP for HPC] to run inference locally.
   • Any cloud provider is discouraged to mitigate leakage-risk of PII (sensive and private data).
5. Run preprocessing script in folder w/ your exported data (zip/json). Standardized format, merge, dedup, chunk, and embed.
6. Run batch survey script w/ flags for your exported data, survey questions, and LLM model of choice.
7. Run analysis script to calculate fidelity and plot results.

---

If you are a donor here is the upload and survey link. We'll give you a specific donor ID to add to your url, which only you should use.

Export data from each platform

Export data from each platform

REQUIRED TO DO MESSENGER BEFORE FACEBOOK DUE TO ENCRYPTION-QUIRK!

1. Messenger --> click on your profile picture in bottom left corner --> Privacy & safety --> End-to-end encrypted chats --> Message storage (ensure Chat history Last backup is today's date) --> Download message storage data --> Select All time and lower-quality media --> Download file.
2. Facebook (IG is also exported here in this url) --> Create export --> pick account[s] --> export to device --> Customize information --> Clear all other than Messages (Get All time, in JSON format, and pick lower media quality). Wait for email confirming the export is ready (~3 days for "all time") --> download zip file.
3. Instagram Exact same process as Facebook, just pick your Instagram account instead.
4. WhatsApp (via phone app) --> Settings --> Chats --> Export chat --> pick your 1-on-1 chats to export (instant).
5. Optional: Use Beeper's API to continuosly export new messages. But be aware of our experiment's scope being a snapshot in time.

Demonstrations

1	2	3	4

Note

This fork is a refactored version of this repository, where our original commit history is preserved. We refactored for speed, transparency, and reproducability.