Spaces:
Sleeping
Sleeping
Upload app and README
Browse files
README.md
ADDED
|
@@ -0,0 +1,127 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
title: User Modeling Agent
|
| 3 |
+
emoji: π
|
| 4 |
+
colorFrom: green
|
| 5 |
+
colorTo: red
|
| 6 |
+
sdk: docker
|
| 7 |
+
app_port: 7860
|
| 8 |
+
pinned: false
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# User Modeling Agent
|
| 12 |
+
|
| 13 |
+
**DSN Γ BCT LLM Agent Challenge 2026 β Task A.**
|
| 14 |
+
|
| 15 |
+
An agent that reads a person into a behavioural *persona*, then writes the
|
| 16 |
+
star rating and the review that person would leave for an unseen product β
|
| 17 |
+
and critiques and revises its own draft before returning it.
|
| 18 |
+
|
| 19 |
+
> Live demo: *(your HuggingFace Space URL)*
|
| 20 |
+
> Code: *(your GitHub repo URL)*
|
| 21 |
+
|
| 22 |
+
---
|
| 23 |
+
|
| 24 |
+
## What it does
|
| 25 |
+
|
| 26 |
+
Given a **user persona** and **product details**, the agent produces:
|
| 27 |
+
|
| 28 |
+
- a **star rating** (1β5) the user would likely give, and
|
| 29 |
+
- a **written review** in that user's voice β tone, length, and quirks matched.
|
| 30 |
+
|
| 31 |
+
It is not a generic review generator. Every output is conditioned on a
|
| 32 |
+
specific reader, and the rating is reasoned, not guessed.
|
| 33 |
+
|
| 34 |
+
## The agentic workflow
|
| 35 |
+
|
| 36 |
+
The system is an agent, not a single prompt. It runs a five-step loop:
|
| 37 |
+
|
| 38 |
+
1. **Build the persona.** A `PersonaEngine` extracts a structured persona β
|
| 39 |
+
quantitative signals (average rating, rating spread, review length,
|
| 40 |
+
domains, rating distribution) and qualitative voice (tone, preferred
|
| 41 |
+
themes, common complaints, a one-line voice descriptor) distilled by an
|
| 42 |
+
LLM from sample reviews. In the deployed app the persona can also be
|
| 43 |
+
*composed directly* from typed input β the brief's persona-as-input
|
| 44 |
+
contract.
|
| 45 |
+
|
| 46 |
+
2. **Select grounding history.** For a real user, the agent picks the few
|
| 47 |
+
past reviews most similar to the target item, so it writes from concrete
|
| 48 |
+
evidence of how this person actually phrases things.
|
| 49 |
+
|
| 50 |
+
3. **Generate the rating and review.** A single LLM call, with the rating
|
| 51 |
+
reasoned in two explicit steps β first the persona *prior* (what this
|
| 52 |
+
user usually gives), then the *item evidence* (what the title and
|
| 53 |
+
description signal). The final rating is the prior adjusted by the
|
| 54 |
+
evidence, so a generous reviewer still rates a poor item low and a
|
| 55 |
+
critical reviewer still rates a strong item high.
|
| 56 |
+
|
| 57 |
+
4. **Self-reflection β critique and revise.** A critic LLM audits the draft
|
| 58 |
+
for ratingβtext consistency, voice match, and on-topic fit. If it
|
| 59 |
+
objects, the agent rewrites with that feedback and re-checks β up to two
|
| 60 |
+
cycles. This act β critique β revise loop is what makes it an agent.
|
| 61 |
+
|
| 62 |
+
5. **Post-process.** The rating is clamped to range. An optional Nigerian
|
| 63 |
+
Pidgin rendering layer can restyle the review while preserving meaning,
|
| 64 |
+
sentiment, and rating.
|
| 65 |
+
|
| 66 |
+
The agent degrades gracefully: if an LLM call fails, it falls back to a
|
| 67 |
+
deterministic persona rather than crashing.
|
| 68 |
+
|
| 69 |
+
## How it maps to the Task A rubric
|
| 70 |
+
|
| 71 |
+
- **Review Text Quality** β reviews are grounded in the user's real past
|
| 72 |
+
reviews and self-critiqued for voice match.
|
| 73 |
+
- **Rating Accuracy** β the two-step prior-plus-evidence rating logic
|
| 74 |
+
corrects the common failure of predicting from the user average alone.
|
| 75 |
+
- **Behavioural Fidelity** β persona-conditioned generation; the persona
|
| 76 |
+
portrait is visible in the app for inspection.
|
| 77 |
+
- **Nigerian contextualization (bonus)** β a toggleable Nigerian Pidgin
|
| 78 |
+
rendering layer; off by default so scored output stays standard English.
|
| 79 |
+
|
| 80 |
+
## Running locally
|
| 81 |
+
|
| 82 |
+
```bash
|
| 83 |
+
pip install -r requirements.txt
|
| 84 |
+
# set your key in a .env file: LLM_PROVIDER=gemini and GEMINI_API_KEY=...
|
| 85 |
+
streamlit run app.py
|
| 86 |
+
```
|
| 87 |
+
|
| 88 |
+
The processed data (`data/processed/*.parquet`) must be present.
|
| 89 |
+
|
| 90 |
+
A FastAPI service is also available:
|
| 91 |
+
|
| 92 |
+
```bash
|
| 93 |
+
uvicorn task_a_user_modeling.main:app --reload
|
| 94 |
+
```
|
| 95 |
+
|
| 96 |
+
## Project layout
|
| 97 |
+
|
| 98 |
+
```
|
| 99 |
+
core/ shared engine β config, llm, persona, reflection, nigerian
|
| 100 |
+
task_a_user_modeling/ the Impersonation agent + FastAPI service
|
| 101 |
+
scripts/ test harness (test_task_a.py)
|
| 102 |
+
data/processed/ Amazon Reviews 2023 β Books Β· Movies & TV Β· Kindle Store
|
| 103 |
+
app.py Streamlit demo
|
| 104 |
+
```
|
| 105 |
+
|
| 106 |
+
## Configuration
|
| 107 |
+
|
| 108 |
+
Set in a `.env` file (never commit it):
|
| 109 |
+
|
| 110 |
+
- `LLM_PROVIDER` β `gemini` or `openai`
|
| 111 |
+
- `GEMINI_API_KEY` / `OPENAI_API_KEY`
|
| 112 |
+
|
| 113 |
+
On a HuggingFace Space, set these as **Secrets** in Space settings.
|
| 114 |
+
|
| 115 |
+
## Notes and honest limitations
|
| 116 |
+
|
| 117 |
+
- The self-reflection critic checks internal consistency; it cannot catch a
|
| 118 |
+
rating that is wrong but self-consistent.
|
| 119 |
+
- Rating prediction on hard cases (a critical user who loved something) is
|
| 120 |
+
improved by the two-step logic but can still be ~0.5β1.0β
off.
|
| 121 |
+
- LLM output is non-deterministic; single-run results vary, so evaluation
|
| 122 |
+
averages across many users.
|
| 123 |
+
|
| 124 |
+
## Credits
|
| 125 |
+
|
| 126 |
+
Built for the DSN Γ BCT LLM Agent Challenge 2026.
|
| 127 |
+
Author: *(your name)*. Dataset: Amazon Reviews 2023.
|
app.py
ADDED
|
@@ -0,0 +1,538 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""User Modeling Agent β the demo.
|
| 2 |
+
|
| 3 |
+
DSN Γ BCT LLM Agent Challenge Β· Task A.
|
| 4 |
+
|
| 5 |
+
Takes a user persona and product details as input, and generates a star
|
| 6 |
+
rating and a written review as that user would write it β then critiques
|
| 7 |
+
and revises its own draft (self-reflection). Optionally renders the review
|
| 8 |
+
in Nigerian English.
|
| 9 |
+
|
| 10 |
+
Two ways to use it:
|
| 11 |
+
1. Compose a persona β type a persona + product (the brief's input contract)
|
| 12 |
+
2. Dataset reader β pick a real user, compare against ground truth
|
| 13 |
+
|
| 14 |
+
Run:
|
| 15 |
+
streamlit run app.py
|
| 16 |
+
"""
|
| 17 |
+
from __future__ import annotations
|
| 18 |
+
|
| 19 |
+
import html
|
| 20 |
+
import sys
|
| 21 |
+
from pathlib import Path
|
| 22 |
+
|
| 23 |
+
ROOT = Path(__file__).resolve().parent
|
| 24 |
+
if str(ROOT) not in sys.path:
|
| 25 |
+
sys.path.insert(0, str(ROOT))
|
| 26 |
+
|
| 27 |
+
import pandas as pd
|
| 28 |
+
import streamlit as st
|
| 29 |
+
|
| 30 |
+
from core.config import settings
|
| 31 |
+
from core.persona import PersonaEngine, UserPersona
|
| 32 |
+
from task_a_user_modeling.agent import ImpersonationAgent, ItemInput
|
| 33 |
+
|
| 34 |
+
st.set_page_config(page_title="User Modeling Agent", page_icon="βΆ",
|
| 35 |
+
layout="wide", initial_sidebar_state="expanded")
|
| 36 |
+
|
| 37 |
+
esc = html.escape
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 41 |
+
# Design system
|
| 42 |
+
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 43 |
+
|
| 44 |
+
CSS = """
|
| 45 |
+
<style>
|
| 46 |
+
@import url('https://fonts.googleapis.com/css2?family=Fraunces:opsz,wght@9..144,400;9..144,500;9..144,600;9..144,900&family=Newsreader:ital,opsz,wght@0,6..72,400;0,6..72,500;0,6..72,600;1,6..72,400&family=Spline+Sans+Mono:wght@400;500;600&display=swap');
|
| 47 |
+
|
| 48 |
+
:root {
|
| 49 |
+
--paper:#f3ecdb; --paper-2:#fffdf6; --paper-3:#ece2cb;
|
| 50 |
+
--pine:#1d3a2b; --pine-2:#2c5440; --pine-ink:#14241b;
|
| 51 |
+
--clay:#b0472b; --ochre:#c98a3c; --gold:#d8a64a;
|
| 52 |
+
--ink:#221e16; --muted:#6f6651; --hair:#d4c8aa;
|
| 53 |
+
}
|
| 54 |
+
.stApp { background:var(--paper); color:var(--ink); }
|
| 55 |
+
.stApp::before {
|
| 56 |
+
content:""; position:fixed; inset:0; pointer-events:none; z-index:0;
|
| 57 |
+
background:
|
| 58 |
+
radial-gradient(900px 600px at 12% -5%, rgba(45,84,64,.10), transparent 60%),
|
| 59 |
+
radial-gradient(800px 600px at 95% 8%, rgba(176,71,43,.08), transparent 55%);
|
| 60 |
+
}
|
| 61 |
+
[data-testid="stMainBlockContainer"] { max-width:1140px; padding-top:2rem; padding-bottom:4rem; }
|
| 62 |
+
h1,h2,h3,h4 { font-family:'Fraunces',Georgia,serif !important; color:var(--pine) !important;
|
| 63 |
+
letter-spacing:-0.015em; font-weight:600 !important; }
|
| 64 |
+
html,body,p,div,span,label,li,.stMarkdown { font-family:'Newsreader',Georgia,serif; }
|
| 65 |
+
.stCaption,[data-testid="stCaptionContainer"] { font-family:'Spline Sans Mono',monospace !important; }
|
| 66 |
+
|
| 67 |
+
.masthead { position:relative; z-index:1; margin-bottom:0.3rem; }
|
| 68 |
+
.mast-rule { height:2px; background:var(--pine); margin-bottom:0.5rem; }
|
| 69 |
+
.mast-kicker { font-family:'Spline Sans Mono',monospace; font-size:0.70rem;
|
| 70 |
+
letter-spacing:0.30em; text-transform:uppercase; color:var(--clay); font-weight:600; }
|
| 71 |
+
.mast-title { font-family:'Fraunces',serif; font-weight:900;
|
| 72 |
+
font-size:clamp(2.3rem,5.5vw,3.7rem); line-height:1.0; color:var(--pine);
|
| 73 |
+
margin:0.16rem 0 0.1rem; letter-spacing:-0.03em; }
|
| 74 |
+
.mast-title .em { color:var(--clay); font-style:italic; font-weight:500; }
|
| 75 |
+
.mast-stand { font-family:'Newsreader',serif; font-size:1.08rem; color:#45402f;
|
| 76 |
+
max-width:66ch; line-height:1.45; }
|
| 77 |
+
.mast-stand em { color:var(--clay); font-style:italic; }
|
| 78 |
+
.mast-rule-bot { height:1px; background:var(--hair); margin:0.85rem 0 0.2rem; }
|
| 79 |
+
|
| 80 |
+
.sec-label { font-family:'Spline Sans Mono',monospace; font-size:0.70rem;
|
| 81 |
+
letter-spacing:0.2em; text-transform:uppercase; color:var(--clay);
|
| 82 |
+
font-weight:600; margin:0.3rem 0 0.15rem; }
|
| 83 |
+
|
| 84 |
+
.card { background:var(--paper-2); border:1px solid var(--hair); border-radius:3px;
|
| 85 |
+
padding:1.1rem 1.3rem; margin:0.5rem 0 0.85rem; position:relative; z-index:1; }
|
| 86 |
+
.card-kicker { font-family:'Spline Sans Mono',monospace; font-size:0.64rem;
|
| 87 |
+
letter-spacing:0.2em; text-transform:uppercase; color:var(--clay);
|
| 88 |
+
font-weight:600; margin-bottom:0.5rem; }
|
| 89 |
+
|
| 90 |
+
.persona-quote { font-family:'Fraunces',serif; font-weight:500; font-style:italic;
|
| 91 |
+
font-size:1.26rem; line-height:1.34; color:var(--pine); margin:0.1rem 0 0.8rem;
|
| 92 |
+
padding-left:0.8rem; border-left:3px solid var(--ochre); }
|
| 93 |
+
.pstats { display:flex; gap:1.7rem; flex-wrap:wrap; align-items:flex-end; }
|
| 94 |
+
.pstat .num { font-family:'Fraunces',serif; font-weight:900; font-size:1.5rem;
|
| 95 |
+
color:var(--pine); line-height:1; }
|
| 96 |
+
.pstat .lab { font-family:'Spline Sans Mono',monospace; font-size:0.60rem;
|
| 97 |
+
letter-spacing:0.13em; text-transform:uppercase; color:var(--muted); margin-top:0.2rem; }
|
| 98 |
+
.chips { margin-top:0.6rem; }
|
| 99 |
+
.chip-lab { font-family:'Spline Sans Mono',monospace; font-size:0.60rem;
|
| 100 |
+
letter-spacing:0.12em; text-transform:uppercase; color:var(--muted); margin-right:0.4rem; }
|
| 101 |
+
.chip { display:inline-block; margin:0.15rem 0.25rem 0.15rem 0; padding:0.15rem 0.6rem;
|
| 102 |
+
border-radius:999px; font-family:'Spline Sans Mono',monospace; font-size:0.72rem;
|
| 103 |
+
background:var(--paper-3); color:var(--pine-2); border:1px solid var(--hair); }
|
| 104 |
+
.chip.warn { background:#f0ddd2; color:var(--clay); border-color:#e3c4b4; }
|
| 105 |
+
|
| 106 |
+
.panel { background:var(--pine-ink); border-radius:3px; padding:1.35rem 1.55rem;
|
| 107 |
+
margin:0.5rem 0 0.85rem; position:relative; z-index:1;
|
| 108 |
+
box-shadow:0 14px 34px -22px rgba(20,36,27,.7); }
|
| 109 |
+
.panel .card-kicker { color:var(--gold); }
|
| 110 |
+
.rating-row { display:flex; align-items:center; gap:0.8rem; margin:0.25rem 0 0.65rem; }
|
| 111 |
+
.rating-chip { font-family:'Fraunces',serif; font-weight:900; font-size:1.6rem;
|
| 112 |
+
background:var(--clay); color:#fff7ec; padding:0.05rem 0.65rem; border-radius:3px; }
|
| 113 |
+
.stars { font-size:1.15rem; letter-spacing:0.1em; color:var(--gold); }
|
| 114 |
+
.review-body { font-family:'Newsreader',serif; font-size:1.1rem; line-height:1.7;
|
| 115 |
+
color:#f0e9d6; white-space:pre-wrap; }
|
| 116 |
+
.naija-badge { display:inline-block; margin-left:0.45rem; font-family:'Spline Sans Mono',monospace;
|
| 117 |
+
font-size:0.60rem; letter-spacing:0.12em; font-weight:600; background:#e9f0e2;
|
| 118 |
+
color:var(--pine); padding:0.12rem 0.5rem; border-radius:999px; border:1px solid #cdd9bf; }
|
| 119 |
+
|
| 120 |
+
.stepper { display:flex; gap:0; margin:0.3rem 0 0.2rem; flex-wrap:wrap; }
|
| 121 |
+
.step { flex:1; min-width:125px; padding:0.5rem 0.65rem; position:relative; }
|
| 122 |
+
.step .dot { width:11px; height:11px; border-radius:50%; background:var(--pine); margin-bottom:0.35rem; }
|
| 123 |
+
.step.flag .dot { background:var(--clay); }
|
| 124 |
+
.step.pass .dot { background:var(--pine-2); }
|
| 125 |
+
.step .st-name { font-family:'Fraunces',serif; font-weight:600; font-size:0.93rem;
|
| 126 |
+
color:var(--pine); line-height:1.1; }
|
| 127 |
+
.step .st-sub { font-family:'Spline Sans Mono',monospace; font-size:0.63rem;
|
| 128 |
+
color:var(--muted); margin-top:0.18rem; }
|
| 129 |
+
.step:not(:last-child)::after { content:""; position:absolute; top:0.87rem; right:-2px;
|
| 130 |
+
width:100%; height:1px;
|
| 131 |
+
background:repeating-linear-gradient(90deg,var(--hair) 0 6px,transparent 6px 12px); }
|
| 132 |
+
.critique-note { font-family:'Newsreader',serif; font-style:italic; font-size:0.93rem;
|
| 133 |
+
color:#5a4030; line-height:1.45; background:#f0ddd2; border-left:3px solid var(--clay);
|
| 134 |
+
padding:0.5rem 0.75rem; border-radius:2px; margin-top:0.45rem; }
|
| 135 |
+
|
| 136 |
+
.cmp { background:var(--paper-2); border:1px solid var(--hair); border-radius:3px;
|
| 137 |
+
padding:0.9rem 1.05rem; height:100%; }
|
| 138 |
+
.cmp.truth { border-top:3px solid var(--pine-2); }
|
| 139 |
+
.cmp.agent { border-top:3px solid var(--clay); }
|
| 140 |
+
.cmp-head { font-family:'Spline Sans Mono',monospace; font-size:0.62rem;
|
| 141 |
+
letter-spacing:0.15em; text-transform:uppercase; color:var(--muted); margin-bottom:0.35rem; }
|
| 142 |
+
.cmp-body { font-family:'Newsreader',serif; font-size:0.97rem; line-height:1.5;
|
| 143 |
+
color:#4a4434; white-space:pre-wrap; }
|
| 144 |
+
.delta { font-family:'Spline Sans Mono',monospace; font-size:0.70rem; font-weight:600;
|
| 145 |
+
padding:0.16rem 0.55rem; border-radius:999px; }
|
| 146 |
+
.delta.good { background:#e3ecd9; color:var(--pine); }
|
| 147 |
+
.delta.mid { background:#f3e6c8; color:#8a6420; }
|
| 148 |
+
.delta.far { background:#f0d8cc; color:var(--clay); }
|
| 149 |
+
|
| 150 |
+
.empty { border:1px dashed var(--hair); border-radius:3px; padding:1.5rem; text-align:center;
|
| 151 |
+
font-family:'Newsreader',serif; font-style:italic; color:var(--muted); font-size:1rem;
|
| 152 |
+
background:rgba(255,253,246,.5); }
|
| 153 |
+
|
| 154 |
+
@keyframes rise { from{opacity:0;transform:translateY(13px);} to{opacity:1;transform:translateY(0);} }
|
| 155 |
+
.reveal { animation:rise 0.55s cubic-bezier(.2,.7,.2,1) both; }
|
| 156 |
+
.d1{animation-delay:.04s;} .d2{animation-delay:.13s;} .d3{animation-delay:.22s;}
|
| 157 |
+
|
| 158 |
+
.stButton > button { background:var(--pine); color:var(--paper); border:none; border-radius:3px;
|
| 159 |
+
font-family:'Spline Sans Mono',monospace; font-weight:600; font-size:0.82rem;
|
| 160 |
+
letter-spacing:0.05em; padding:0.55rem 1rem; }
|
| 161 |
+
.stButton > button:hover { background:var(--clay); color:#fff7ec; }
|
| 162 |
+
[data-testid="stSidebar"] { background:var(--pine-ink); border-right:1px solid #2c4133; }
|
| 163 |
+
[data-testid="stSidebar"] * { color:#e7e0cd; }
|
| 164 |
+
[data-testid="stSidebar"] h1,[data-testid="stSidebar"] h2,[data-testid="stSidebar"] h3 { color:#f3ecdb !important; }
|
| 165 |
+
[data-baseweb="tab-list"] { gap:0.3rem; border-bottom:2px solid var(--pine); }
|
| 166 |
+
[data-baseweb="tab"] { font-family:'Fraunces',serif !important; font-weight:600;
|
| 167 |
+
font-size:1rem; color:var(--muted); }
|
| 168 |
+
[data-baseweb="tab"][aria-selected="true"] { color:var(--pine) !important; }
|
| 169 |
+
[data-baseweb="tab-highlight"] { background:var(--clay) !important; height:3px; }
|
| 170 |
+
.foot { margin-top:2.2rem; padding-top:0.85rem; border-top:1px solid var(--hair);
|
| 171 |
+
font-family:'Spline Sans Mono',monospace; font-size:0.68rem; color:var(--muted); line-height:1.6; }
|
| 172 |
+
</style>
|
| 173 |
+
"""
|
| 174 |
+
st.markdown(CSS, unsafe_allow_html=True)
|
| 175 |
+
|
| 176 |
+
|
| 177 |
+
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 178 |
+
# HTML builders
|
| 179 |
+
# βββββββββββββββοΏ½οΏ½οΏ½ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 180 |
+
|
| 181 |
+
def stars(r: float) -> str:
|
| 182 |
+
f = int(round(r))
|
| 183 |
+
return "β
" * f + "β" * (5 - f)
|
| 184 |
+
|
| 185 |
+
|
| 186 |
+
def persona_card(p: UserPersona) -> str:
|
| 187 |
+
themes = "".join(f'<span class="chip">{esc(t)}</span>'
|
| 188 |
+
for t in p.preferred_themes) or '<span class="chip">β</span>'
|
| 189 |
+
comps = "".join(f'<span class="chip warn">{esc(t)}</span>'
|
| 190 |
+
for t in p.common_complaints) or '<span class="chip warn">β</span>'
|
| 191 |
+
nrev = (f'{p.n_reviews}' if p.n_reviews else 'composed')
|
| 192 |
+
return f"""
|
| 193 |
+
<div class="card reveal d1">
|
| 194 |
+
<div class="card-kicker">The Reader Β· persona</div>
|
| 195 |
+
<div class="persona-quote">β{esc(p.voice_one_liner or 'No voice captured.')}β</div>
|
| 196 |
+
<div class="pstats">
|
| 197 |
+
<div class="pstat"><div class="num">{nrev}</div><div class="lab">history</div></div>
|
| 198 |
+
<div class="pstat"><div class="num">{p.avg_rating:.1f}β
</div><div class="lab">avg rating</div></div>
|
| 199 |
+
<div class="pstat"><div class="num">{esc(p.tone or 'β')}</div><div class="lab">tone</div></div>
|
| 200 |
+
</div>
|
| 201 |
+
<div class="chips"><span class="chip-lab">drawn to</span>{themes}</div>
|
| 202 |
+
<div class="chips"><span class="chip-lab">put off by</span>{comps}</div>
|
| 203 |
+
</div>"""
|
| 204 |
+
|
| 205 |
+
|
| 206 |
+
def reflection_stepper(iters: int, refined: bool, notes: list[str] | None) -> str:
|
| 207 |
+
steps = ['<div class="step pass"><div class="dot"></div>'
|
| 208 |
+
'<div class="st-name">First draft</div>'
|
| 209 |
+
'<div class="st-sub">generated in-voice</div></div>']
|
| 210 |
+
if refined:
|
| 211 |
+
steps += ['<div class="step flag"><div class="dot"></div>'
|
| 212 |
+
'<div class="st-name">Self-critique</div>'
|
| 213 |
+
'<div class="st-sub">found issues</div></div>',
|
| 214 |
+
'<div class="step pass"><div class="dot"></div>'
|
| 215 |
+
'<div class="st-name">Revised draft</div>'
|
| 216 |
+
'<div class="st-sub">rewritten with feedback</div></div>',
|
| 217 |
+
'<div class="step pass"><div class="dot"></div>'
|
| 218 |
+
'<div class="st-name">Re-checked</div>'
|
| 219 |
+
'<div class="st-sub">critique cleared</div></div>']
|
| 220 |
+
else:
|
| 221 |
+
steps += ['<div class="step pass"><div class="dot"></div>'
|
| 222 |
+
'<div class="st-name">Self-critique</div>'
|
| 223 |
+
'<div class="st-sub">passed first pass</div></div>',
|
| 224 |
+
'<div class="step pass"><div class="dot"></div>'
|
| 225 |
+
'<div class="st-name">Accepted</div>'
|
| 226 |
+
'<div class="st-sub">no revision needed</div></div>']
|
| 227 |
+
note = ""
|
| 228 |
+
if notes:
|
| 229 |
+
real = [n for n in notes if n and n.strip().lower() != "passed"]
|
| 230 |
+
if real:
|
| 231 |
+
note = f'<div class="critique-note">The critic flagged: {esc(real[0])}</div>'
|
| 232 |
+
return f"""
|
| 233 |
+
<div class="card reveal d3">
|
| 234 |
+
<div class="card-kicker">Self-reflection Β· {iters} critique cycle(s)</div>
|
| 235 |
+
<div class="stepper">{''.join(steps)}</div>
|
| 236 |
+
{note}
|
| 237 |
+
</div>"""
|
| 238 |
+
|
| 239 |
+
|
| 240 |
+
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 241 |
+
# Cached resources
|
| 242 |
+
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 243 |
+
|
| 244 |
+
@st.cache_data(show_spinner=False)
|
| 245 |
+
def load_data():
|
| 246 |
+
rev = pd.read_parquet(settings.processed_dir / "reviews.parquet")
|
| 247 |
+
items = pd.read_parquet(settings.processed_dir / "items.parquet")
|
| 248 |
+
return rev, items
|
| 249 |
+
|
| 250 |
+
|
| 251 |
+
@st.cache_resource(show_spinner=False)
|
| 252 |
+
def get_engines():
|
| 253 |
+
return PersonaEngine(), ImpersonationAgent()
|
| 254 |
+
|
| 255 |
+
|
| 256 |
+
def composed_persona(desc: str, themes: list[str], dislikes: list[str],
|
| 257 |
+
tone: str, avg_rating: float) -> UserPersona:
|
| 258 |
+
"""Build a UserPersona from typed input β the brief's persona-as-input contract."""
|
| 259 |
+
# rating distribution skewed around the stated average
|
| 260 |
+
lo, hi = int(avg_rating), min(5, int(avg_rating) + 1)
|
| 261 |
+
dist = {lo: 0.55, hi: 0.35} if lo != hi else {lo: 0.9}
|
| 262 |
+
dist.setdefault(3, 0.1)
|
| 263 |
+
return UserPersona(
|
| 264 |
+
user_id="composed", n_reviews=0, avg_rating=avg_rating,
|
| 265 |
+
std_rating=0.6, avg_review_length=90.0, std_review_length=30.0,
|
| 266 |
+
verified_rate=1.0, domains=[], n_domains=0,
|
| 267 |
+
rating_distribution=dist, top_terms=[],
|
| 268 |
+
tone=tone, preferred_themes=themes, common_complaints=dislikes,
|
| 269 |
+
voice_one_liner=desc, history_samples=[],
|
| 270 |
+
)
|
| 271 |
+
|
| 272 |
+
|
| 273 |
+
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 274 |
+
# Masthead
|
| 275 |
+
# βββββββββοΏ½οΏ½ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 276 |
+
|
| 277 |
+
st.markdown("""
|
| 278 |
+
<div class="masthead">
|
| 279 |
+
<div class="mast-rule"></div>
|
| 280 |
+
<div class="mast-kicker">DSN Γ BCT LLM Agent Challenge Β· Task A</div>
|
| 281 |
+
<div class="mast-title">User Modeling <span class="em">Agent</span></div>
|
| 282 |
+
<div class="mast-stand">
|
| 283 |
+
Give it a <em>user persona</em> and a <em>product</em>. It writes the star
|
| 284 |
+
rating and the review that user would write β weighing what they usually do
|
| 285 |
+
against what this specific item signals β then <em>critiques and revises</em>
|
| 286 |
+
its own draft before showing it.
|
| 287 |
+
</div>
|
| 288 |
+
<div class="mast-rule-bot"></div>
|
| 289 |
+
</div>
|
| 290 |
+
""", unsafe_allow_html=True)
|
| 291 |
+
|
| 292 |
+
try:
|
| 293 |
+
reviews, items = load_data()
|
| 294 |
+
except Exception as e:
|
| 295 |
+
st.error(f"Could not load data β ensure data/processed/*.parquet exist.\n\n{e}")
|
| 296 |
+
st.stop()
|
| 297 |
+
|
| 298 |
+
train = reviews[reviews["split"] == "train"]
|
| 299 |
+
test = reviews[reviews["split"] == "test"]
|
| 300 |
+
persona_engine, agent = get_engines()
|
| 301 |
+
|
| 302 |
+
with st.sidebar:
|
| 303 |
+
st.markdown("## βΆ Controls")
|
| 304 |
+
naija = st.toggle("π³π¬ Naija mode", value=False,
|
| 305 |
+
help="Render the review in Nigerian English. Meaning, "
|
| 306 |
+
"sentiment and rating are preserved β only voice shifts.")
|
| 307 |
+
st.caption("Naija mode ON β review in Nigerian English."
|
| 308 |
+
if naija else "Standard English output.")
|
| 309 |
+
st.divider()
|
| 310 |
+
st.markdown("### How it works")
|
| 311 |
+
st.caption("The agent builds a persona, drafts a review in that voice, then "
|
| 312 |
+
"runs a self-reflection loop β a critic LLM checks rating-text "
|
| 313 |
+
"consistency, voice match and on-topic fit, and the agent revises "
|
| 314 |
+
"if the critic objects.")
|
| 315 |
+
st.divider()
|
| 316 |
+
st.caption(f"Built by Israel")
|
| 317 |
+
|
| 318 |
+
st.session_state.setdefault("result", None)
|
| 319 |
+
st.session_state.setdefault("ctx", None)
|
| 320 |
+
|
| 321 |
+
|
| 322 |
+
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 323 |
+
# Tabs β Compose (primary) Β· Dataset reader (secondary)
|
| 324 |
+
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 325 |
+
|
| 326 |
+
tab_compose, tab_dataset = st.tabs(["β Compose a persona",
|
| 327 |
+
"β Dataset reader"])
|
| 328 |
+
|
| 329 |
+
# ββ COMPOSE βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 330 |
+
with tab_compose:
|
| 331 |
+
st.markdown('<div class="sec-label">Input Β· persona and product</div>',
|
| 332 |
+
unsafe_allow_html=True)
|
| 333 |
+
st.markdown("Describe a reader and a product. The agent will write the "
|
| 334 |
+
"review that reader would leave.")
|
| 335 |
+
|
| 336 |
+
cL, cR = st.columns(2)
|
| 337 |
+
with cL:
|
| 338 |
+
st.markdown("**The reader**")
|
| 339 |
+
p_desc = st.text_area(
|
| 340 |
+
"Describe the reader's reviewing voice",
|
| 341 |
+
value="A thoughtful reader who loves character-driven stories and "
|
| 342 |
+
"rich world-building, but is impatient with slow pacing.",
|
| 343 |
+
height=90, key="p_desc")
|
| 344 |
+
p_themes = st.text_input("Drawn to (comma-separated)",
|
| 345 |
+
value="character development, immersive worlds, "
|
| 346 |
+
"original plots", key="p_themes")
|
| 347 |
+
p_dislikes = st.text_input("Put off by (comma-separated)",
|
| 348 |
+
value="slow pacing, thin characters", key="p_dis")
|
| 349 |
+
c1, c2 = st.columns(2)
|
| 350 |
+
with c1:
|
| 351 |
+
p_tone = st.selectbox("Tone", ["enthusiastic", "analytical", "casual",
|
| 352 |
+
"critical", "earnest", "terse"], key="p_tone")
|
| 353 |
+
with c2:
|
| 354 |
+
p_rating = st.slider("Typical rating", 1.0, 5.0, 4.0, 0.5, key="p_rate")
|
| 355 |
+
with cR:
|
| 356 |
+
st.markdown("**The product**")
|
| 357 |
+
i_title = st.text_input("Title",
|
| 358 |
+
value="The Midnight Library", key="i_title")
|
| 359 |
+
i_domain = st.selectbox("Domain", ["Books", "Movies_and_TV", "Kindle_Store"],
|
| 360 |
+
key="i_domain")
|
| 361 |
+
i_desc = st.text_area(
|
| 362 |
+
"Description / synopsis",
|
| 363 |
+
value="A novel about a library between life and death, where each "
|
| 364 |
+
"book lets a woman try a different version of her life.",
|
| 365 |
+
height=110, key="i_desc")
|
| 366 |
+
|
| 367 |
+
go = st.button("Generate review βΆ", key="go_compose", use_container_width=True)
|
| 368 |
+
|
| 369 |
+
if go:
|
| 370 |
+
try:
|
| 371 |
+
with st.status("The agent is workingβ¦", expanded=True) as status:
|
| 372 |
+
themes = [t.strip() for t in p_themes.split(",") if t.strip()]
|
| 373 |
+
dislikes = [t.strip() for t in p_dislikes.split(",") if t.strip()]
|
| 374 |
+
st.write("Assembling the personaβ¦")
|
| 375 |
+
persona = composed_persona(p_desc, themes, dislikes, p_tone, p_rating)
|
| 376 |
+
item = ItemInput(parent_asin="composed", title=i_title,
|
| 377 |
+
description=i_desc, categories="",
|
| 378 |
+
domain=i_domain)
|
| 379 |
+
st.write("Drafting in the reader's voice, then self-critiquingβ¦")
|
| 380 |
+
result = agent.run(persona, item, naija_mode=naija)
|
| 381 |
+
st.write("Self-reflection complete")
|
| 382 |
+
status.update(label="Review generated", state="complete")
|
| 383 |
+
st.session_state.result = result
|
| 384 |
+
st.session_state.ctx = {"persona": persona, "item": item, "truth": None}
|
| 385 |
+
except Exception as e:
|
| 386 |
+
st.session_state.result = None
|
| 387 |
+
st.markdown(f'<div class="card" style="border-left:3px solid var(--clay)">'
|
| 388 |
+
f'<div class="card-kicker">Generation interrupted</div>'
|
| 389 |
+
f'The model call did not complete β it may be rate-limited. '
|
| 390 |
+
f'Try again shortly.<br><span style="font-family:Spline Sans Mono,'
|
| 391 |
+
f'monospace;font-size:0.72rem;color:#6f6651">'
|
| 392 |
+
f'{esc(type(e).__name__)}</span></div>', unsafe_allow_html=True)
|
| 393 |
+
|
| 394 |
+
# ββ DATASET READER ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 395 |
+
with tab_dataset:
|
| 396 |
+
st.markdown('<div class="sec-label">Input Β· a real reader from the data</div>',
|
| 397 |
+
unsafe_allow_html=True)
|
| 398 |
+
st.markdown("Pick a reader. The agent builds their persona from real history "
|
| 399 |
+
"and writes a review of a held-out item β compared to what they "
|
| 400 |
+
"actually wrote.")
|
| 401 |
+
|
| 402 |
+
elig = train.groupby("user_id").size().reset_index(name="n")
|
| 403 |
+
elig = elig[(elig["n"] >= 5) & (elig["user_id"].isin(set(test["user_id"])))]
|
| 404 |
+
users = elig.sample(min(40, len(elig)), random_state=7)["user_id"].tolist()
|
| 405 |
+
|
| 406 |
+
cc1, cc2 = st.columns([3, 1])
|
| 407 |
+
with cc1:
|
| 408 |
+
user = st.selectbox("Reader", users, key="sel_user",
|
| 409 |
+
label_visibility="collapsed")
|
| 410 |
+
with cc2:
|
| 411 |
+
go_ds = st.button("Generate βΆ", key="go_ds", use_container_width=True)
|
| 412 |
+
|
| 413 |
+
if go_ds and user:
|
| 414 |
+
try:
|
| 415 |
+
with st.status("The agent is workingβ¦", expanded=True) as status:
|
| 416 |
+
ut = test[test["user_id"] == user]
|
| 417 |
+
if ut.empty:
|
| 418 |
+
status.update(label="No held-out item for this reader",
|
| 419 |
+
state="error")
|
| 420 |
+
st.stop()
|
| 421 |
+
tr = ut.iloc[0]
|
| 422 |
+
tid = tr["parent_asin"]
|
| 423 |
+
meta = items[items["parent_asin"] == tid]
|
| 424 |
+
if meta.empty:
|
| 425 |
+
item = ItemInput(parent_asin=tid, title=str(tr.get("title", "")),
|
| 426 |
+
description="", categories="", domain=tr["domain"])
|
| 427 |
+
else:
|
| 428 |
+
m = meta.iloc[0]
|
| 429 |
+
item = ItemInput(parent_asin=tid, title=str(m.get("title", "")),
|
| 430 |
+
description=str(m.get("description", ""))[:1500],
|
| 431 |
+
categories=str(m.get("categories", "")),
|
| 432 |
+
domain=tr["domain"],
|
| 433 |
+
average_rating=(float(m["average_rating"])
|
| 434 |
+
if pd.notna(m.get("average_rating"))
|
| 435 |
+
else None))
|
| 436 |
+
st.write("Reading the reader's historyβ¦")
|
| 437 |
+
persona = persona_engine.from_dataframe(user, train)
|
| 438 |
+
persona = persona_engine.enrich(persona)
|
| 439 |
+
st.write(f"Persona built from {persona.n_reviews} reviews")
|
| 440 |
+
st.write("Drafting in their voice, then self-critiquingβ¦")
|
| 441 |
+
result = agent.run(persona, item, naija_mode=naija)
|
| 442 |
+
st.write("Self-reflection complete")
|
| 443 |
+
status.update(label="Review generated", state="complete")
|
| 444 |
+
st.session_state.result = result
|
| 445 |
+
st.session_state.ctx = {"persona": persona, "item": item,
|
| 446 |
+
"truth": {"rating": float(tr["rating"]),
|
| 447 |
+
"text": str(tr["text"])}}
|
| 448 |
+
except Exception as e:
|
| 449 |
+
st.session_state.result = None
|
| 450 |
+
st.markdown(f'<div class="card" style="border-left:3px solid var(--clay)">'
|
| 451 |
+
f'<div class="card-kicker">Generation interrupted</div>'
|
| 452 |
+
f'The model call did not complete β it may be rate-limited. '
|
| 453 |
+
f'Try again shortly.<br><span style="font-family:Spline Sans Mono,'
|
| 454 |
+
f'monospace;font-size:0.72rem;color:#6f6651">'
|
| 455 |
+
f'{esc(type(e).__name__)}</span></div>', unsafe_allow_html=True)
|
| 456 |
+
|
| 457 |
+
|
| 458 |
+
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 459 |
+
# Result β shown below both tabs
|
| 460 |
+
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 461 |
+
|
| 462 |
+
res = st.session_state.result
|
| 463 |
+
ctx = st.session_state.ctx
|
| 464 |
+
st.markdown("---")
|
| 465 |
+
if res and ctx:
|
| 466 |
+
st.markdown(persona_card(ctx["persona"]), unsafe_allow_html=True)
|
| 467 |
+
|
| 468 |
+
it = ctx["item"]
|
| 469 |
+
st.markdown(f"""
|
| 470 |
+
<div class="card reveal d2">
|
| 471 |
+
<div class="card-kicker">The Item</div>
|
| 472 |
+
<span style="font-family:Spline Sans Mono,monospace;font-size:0.6rem;
|
| 473 |
+
letter-spacing:0.13em;text-transform:uppercase;color:var(--pine-2)">
|
| 474 |
+
{esc(it.domain)}</span>
|
| 475 |
+
<div style="font-family:Fraunces,serif;font-weight:600;font-size:1.14rem;
|
| 476 |
+
color:var(--ink);margin-top:0.1rem">{esc(it.title)}</div>
|
| 477 |
+
</div>""", unsafe_allow_html=True)
|
| 478 |
+
|
| 479 |
+
badge = '<span class="naija-badge">NAIJA VOICE</span>' if res.naija_mode else ""
|
| 480 |
+
st.markdown(f"""
|
| 481 |
+
<div class="panel reveal d3">
|
| 482 |
+
<div class="card-kicker">The Generated Review Β· written as the reader</div>
|
| 483 |
+
<div class="rating-row">
|
| 484 |
+
<span class="rating-chip">{res.rating:.1f}</span>
|
| 485 |
+
<span class="stars">{stars(res.rating)}</span>{badge}
|
| 486 |
+
</div>
|
| 487 |
+
<div class="review-body">{esc(res.review)}</div>
|
| 488 |
+
</div>""", unsafe_allow_html=True)
|
| 489 |
+
|
| 490 |
+
st.markdown(reflection_stepper(res.reflection_iterations,
|
| 491 |
+
res.reflection_refined,
|
| 492 |
+
res.reflection_notes), unsafe_allow_html=True)
|
| 493 |
+
|
| 494 |
+
st.markdown('<div class="sec-label">Why this rating</div>', unsafe_allow_html=True)
|
| 495 |
+
truth = ctx.get("truth")
|
| 496 |
+
if truth:
|
| 497 |
+
col1, col2 = st.columns(2)
|
| 498 |
+
with col1:
|
| 499 |
+
st.markdown(f"""
|
| 500 |
+
<div class="cmp agent reveal d1">
|
| 501 |
+
<div class="cmp-head">The agent rated it {res.rating:.1f}β
</div>
|
| 502 |
+
<div class="cmp-body">{esc(res.reasoning)}</div>
|
| 503 |
+
</div>""", unsafe_allow_html=True)
|
| 504 |
+
with col2:
|
| 505 |
+
d = abs(res.rating - truth["rating"])
|
| 506 |
+
dc = "good" if d <= 0.5 else ("mid" if d <= 1.0 else "far")
|
| 507 |
+
t = truth["text"].replace("<br />", "\n").replace("<br>", "\n")
|
| 508 |
+
t = t[:520] + ("β¦" if len(t) > 520 else "")
|
| 509 |
+
st.markdown(f"""
|
| 510 |
+
<div class="cmp truth reveal d2">
|
| 511 |
+
<div class="cmp-head">The reader actually wrote
|
| 512 |
+
<span class="delta {dc}">Ξ {d:.1f}β
</span></div>
|
| 513 |
+
<div style="margin:0.15rem 0 0.35rem">
|
| 514 |
+
<span class="stars" style="color:var(--pine-2)">{stars(truth['rating'])}</span>
|
| 515 |
+
<span style="font-family:Spline Sans Mono,monospace;font-size:0.74rem;
|
| 516 |
+
color:#6f6651"> {truth['rating']:.1f}β
</span></div>
|
| 517 |
+
<div class="cmp-body">{esc(t)}</div>
|
| 518 |
+
</div>""", unsafe_allow_html=True)
|
| 519 |
+
else:
|
| 520 |
+
st.markdown(f"""
|
| 521 |
+
<div class="cmp agent reveal d1">
|
| 522 |
+
<div class="cmp-head">The agent rated it {res.rating:.1f}β
</div>
|
| 523 |
+
<div class="cmp-body">{esc(res.reasoning)}</div>
|
| 524 |
+
</div>""", unsafe_allow_html=True)
|
| 525 |
+
st.caption(f"grounded on {res.used_history_count} similar past reviews")
|
| 526 |
+
else:
|
| 527 |
+
st.markdown('<div class="empty">Compose a persona and a product, or pick a '
|
| 528 |
+
'dataset reader β then press <b>Generate</b>. The agent writes '
|
| 529 |
+
'the review in that reader\'s voice and shows its reasoning.</div>',
|
| 530 |
+
unsafe_allow_html=True)
|
| 531 |
+
|
| 532 |
+
st.markdown("""
|
| 533 |
+
<div class="foot">
|
| 534 |
+
User Modeling Agent Β· DSN Γ BCT LLM Agent Challenge 2026 Β·
|
| 535 |
+
persona β draft in-voice β self-reflection critique & revise Β·
|
| 536 |
+
rating predicted as persona prior adjusted by item evidence
|
| 537 |
+
</div>
|
| 538 |
+
""", unsafe_allow_html=True)
|