README: AgiHouse Hackathon - Putin Persona Simulation (RAG Implementation)

Project Goal:

This project explores the potential of using Large Language Models (LLMs) combined with Retrieval-Augmented Generation (RAG) to simulate the strategic thinking and decision-making processes of world leaders, specifically focusing on Vladimir Putin. The objective is to create an AI-driven chatbot persona that models Putin's inferred private rationale and unfiltered thoughts, distinct from his public statements, to better understand potential future actions and strategic intentions of the Russian Federation.

Architecture Overview:

This project implements a Retrieval-Augmented Generation (RAG) system:

1. Data Ingestion: Text documents (speeches, articles, etc.) and associated metadata are processed, chunked, and embedded using Voyage AI.
2. Storage: File metadata, text chunks, and their vector embeddings are stored in a Supabase PostgreSQL database (likely utilizing pgvector).
3. Query Processing: When a user sends a message:
   • The query is embedded using Voyage AI.
   • Supabase is queried to find semantically similar text chunks based on vector similarity.
   • The retrieved chunks are reranked for relevance to the specific query using Voyage AI's reranking API.
4. LLM Interaction: The top reranked chunks are passed as context, along with the user's message and a detailed system prompt defining the "Unfiltered Putin" persona, to OpenAI's GPT-4o model via the BAML framework.
5. Response Streaming: The response generated by GPT-4o is streamed back to the user interface using Server-Sent Events (SSE).

Methodology:

1. Data Aggregation & Processing:

   • We gathered diverse data sources (as described previously: primary, Russian, Western, etc.).
   • Scripts (scripts/process_data.ts, scripts/to_txt.ts) process local text files and associated metadata files.
   • Text content is chunked into manageable segments (services/supabase.ts).
   • Voyage AI (services/voyage.ts) is used to generate vector embeddings for these text chunks (voyage-3 model noted in batch processing).
   • Processed data (file metadata, chunks, embeddings) is stored in Supabase (services/supabase.ts).

2. Retrieval & Reranking:

   • User queries are embedded using Voyage AI (voyage-01 query model noted).
   • A custom Supabase RPC function (query_embeddings) performs vector similarity search (services/query.ts).
   • Retrieved results are reranked for contextual relevance using Voyage AI's reranking model (rerank-2 model noted) to select the best context for the LLM (services/voyage.ts, services/query.ts).

3. Persona Modeling & Generation:

   • BAML (@boundaryml/baml) is used to manage LLM interactions, define prompt structures, and generate typed clients (baml_client/, baml_src/).
   • The core "Unfiltered Putin" system prompt is explicitly defined within baml_src/chat.baml, instructing the AI on persona, tone, and worldview.
   • The BAML Chat function integrates the user message, the reranked context (rerankedResults), and the system prompt.
   • OpenAI's GPT-4o is configured as the underlying LLM (baml_src/clients.baml).

4. API & Interface:

   • An Express.js server (server.ts) written in TypeScript handles API requests.
   • The /chat endpoint orchestrates the RAG process (search, rerank, LLM call) and streams the response back to the client via Server-Sent Events (SSE).
   • An optional /transcribe endpoint utilizes OpenAI Whisper (services/transcription-service.ts) for potential audio input.
   • A static frontend (public/) provides the user interface.

Hypothesis & Potential Applications:

Our hypothesis is that by simulating a leader's inferred private reasoning based on comprehensive data, AI can serve as a valuable tool for geopolitical analysis and prediction.

• Enhanced Understanding: This approach aims to provide deeper insights into the potential motivations and strategic calculus driving Russian actions, distinguishing between public positioning and potential private intent.
• Predictive Analysis Support: While not a replacement for traditional intelligence gathering, this simulation can potentially:
  • Generate plausible scenarios based on the leader's modeled thinking.
  • Help analysts test hypotheses about potential reactions to specific events.
  • Identify potential areas of strategic interest or concern for the Russian Federation based on past patterns and inferred logic.
• Augmenting Intelligence: This tool could augment existing intelligence analysis by providing a dynamic model to explore potential decision pathways, potentially offering early indicators or alternative perspectives on future Russian actions and interests. It allows for exploring "what-if" scenarios based on the simulated persona's logic.

Technology Stack:

• Language/Runtime: TypeScript, Node.js
• Web Framework: Express.js
• LLM Interaction Framework: BAML (@boundaryml/baml)
• LLM: OpenAI GPT-4o
• Vector Embeddings & Reranking: Voyage AI (voyageai library, models: voyage-01, voyage-3, rerank-2)
• Vector Database: Supabase (PostgreSQL with pgvector implied)
• Audio Transcription (Optional): OpenAI Whisper
• (Likely) Frontend: HTML, CSS, JavaScript

Limitations & Considerations:

• Simulation, Not Reality: This is an AI interpretation based on available data, not a factual representation of Vladimir Putin's private thoughts.
• Data Bias: The output is inherently influenced by the biases present in the training data and the selected sources.
• Speculative Nature: Modeling "unfiltered thoughts" is inherently speculative.
• Ethical Sensitivity: Simulating a living world leader involved in significant global events requires careful consideration. This project is undertaken strictly for analytical exploration within the hackathon context.
• Hackathon Scope: This is a proof-of-concept demonstrating the approach, developed within a limited timeframe.

Conclusion:

This project demonstrates a RAG-based application of LLMs for geopolitical analysis. By attempting to model the inferred, unfiltered strategic thinking of a world leader, informed by relevant data retrieved in real-time, we aim to showcase how AI can contribute to a more nuanced understanding of international relations and potentially enhance predictive intelligence capabilities regarding the actions of states like the Russian Federation.