Most chat analyzers give you a word cloud, a sentiment line, and a heatmap.
I wanted something different: a local tool that takes a chat history with
a partner and maps it onto frameworks that relationship scientists have
actually studied for 40+ years.
I spent a weekend building ChatRel.
🚀 Live demo (no install, synthetic data): https://ruanflux.github.io/chatrel/
The 7 frameworks (all with citations)
| Framework | Measures | Source |
|---|---|---|
| Sternberg Triangular | Intimacy × Passion × Commitment, weekly | Sternberg (1986) |
| Knapp Stages | 10-stage relational development | Knapp (1978) |
| ECR-R Attachment | Anxiety × Avoidance, Bartholomew 4-category | Fraley et al. (2000) |
| Gottman 5:1 + Four Horsemen | Positive-negative ratio, criticism/contempt/defensiveness/stonewalling | Gottman (1994) |
| Social Penetration | Self-disclosure depth | Altman & Taylor (1973) |
| Critical Slowing (EWS) | Variance + AR(1) as regime-shift warnings | Scheffer et al. (2009) |
| LIWC (Chinese proxy) | Function-word / cognitive / affect markers | Pennebaker |
What you get
Seven interactive HTML reports:
- Weekly IDS (Intimacy Depth Score) with Sternberg components decomposed
- Attachment 2D scatter placing you and your partner on anxiety × avoidance
- Conflict-repair events with initiator, peak, apology lag, and resolution
- External resilience score (how well the relationship handles poaching / ex-mentions)
- Memory book — auto-curated firsts, sweetest, and deepest moments
- 12-week forecast (base / optimistic / pessimistic scenarios)
Privacy is the architecture, not a feature
- 100% local. No API calls. No telemetry.
- Supports Telegram JSON / generic CSV / WeChat (via external decryption)
- Ships with a synthetic 90-day demo chat — you can evaluate the whole tool without real data
-
.gitignoreblocks every pattern that would leak chat content
Known limits (documented honestly)
- n=1 validation. Nothing calibrated against ECR-R self-report scores. Proxy signals, not diagnoses.
- Forecasts are damped linear extrapolation with mean reversion, not Bayesian.
- Contempt detection had to exclude
笑死(modern Chinese "lmao", not contempt) — similar traps are in-file documented. - Lexicons are Chinese-only. PRs for other languages welcome.
Stack
Python 3.10+ · jieba · snownlp · pandas · plotly. MIT licensed.
Try it
bash
git clone https://github.com/ruanflux/chatrel.git
cd chatrel
pip install -e .
chatrel sample --out sample_data/synthetic_couple.json
chatrel analyze sample_data/synthetic_couple.json --out ./output
