AdolfoJCJ/TIE-Dialog

TIE–Dialog · Conversational Coherence Meter

TIE–Dialog is a Streamlit app that measures conversational coherence using the Theory of Informational Emergence (TIE). It blends semantic embeddings with lexical context to estimate coherence over turns, detects ruptures and recoveries, highlights topic shifts, ranks participants by contribution, and exports an executive narrative (CSV / JSON / PDF).

Works on CPU; CUDA is optional. Exports Unicode-safe PDFs if you include a TTF font (see below).

✨ Features

• —Hybrid coherence C_t SBERT (sentence-transformers) + TF-IDF with robust p10–p90 normalization and soft penalties for noise / low density.

• —Conversation nucleus C_im EMA-smoothed reference field; adaptive threshold Φ(t) and band δ(t) per window.

• —Topic shifts (Novelty) Hysteresis + optional KMeans “mini-clusters”; consolidation to avoid over-flagging.

• —Ruptures & recoveries Strict rules mixing ΔCt, z-score, Cim trends, and topic gating.

• —Per-participant panel & ranking Optionally weighted by number of turns or net impact vs. global mean.

• —Narrative generator Human-style summary of peaks, valleys, topic shifts, ruptures/recoveries, and standout participants.

• —Exports CSV, JSON, and PDF with charts, rankings, and narrative.

• —Presets & controls Smoothing, sensitivity, noise guard, Auto-Calibrate per dialog.

• —Performance CPU-safe mode, batching, caching of embeddings.

• —Optional / legacy Informational node map (networkx) if enabled in your branch.

🧰 Built With

streamlit, sentence-transformers, torch, pandas, numpy, matplotlib, scikit-learn, scipy, plotly, reportlab (PDF), fpdf2 (optional)

🖥️ Installation

Python 3.12 recommended (matches HF Spaces runtime)

python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt