# Athena Oracle — AI Research Intelligence Pipeline

> Multi-source AI news aggregation, scoring, and multi-variant edition rendering.
> One pipeline → multiple audience-specific editions.

**Live site:** https://ai-oracle.com (rendered from this pipeline)

## Architecture

```
┌─────────────────────────────────────────────────────┐
│  adapters/                                          │
│  arxiv  ·  github  ·  hackernews  ·  reddit        │
│  huggingface  ·  rss_feeds                          │
└───────────────┬─────────────────────────────────────┘
                │ fetch(limit, timeout=10s)
                ▼
┌─────────────────────────────────────────────────────┐
│  oracle/ — Core Package                             │
│                                                     │
│  scoring.py    Pure-rule component scoring (0-10)   │
│  dedup.py      Content-hash dedup + verdict engine  │
│  variants.py   Multi-variant edition engine         │
│  render.py     HTML/JSON variant renderer           │
│  summarize.py  Source-aware text summarization      │
│  recency.py    Age-based freshness gate             │
│  themes.py     Theme-based trend tracking           │
│  archive.py    Soft-cap entry archival              │
│  db.py         Schema management + migrations       │
│  config.py     Centralized configuration            │
│  cli.py        Unified CLI (python -m oracle)       │
└───────────────┬─────────────────────────────────────┘
                │
                ▼
┌─────────────────────────────────────────────────────┐
│  oracle.db  — SQLite (784+ entries)                 │
│  Columns: signal_score · final_score · content_hash │
│          · verdict · source_tier · summary           │
└─────────────────────────────────────────────────────┘
                │
                ▼
┌─────────────────────────────────────────────────────┐
│  variants/ — Edition Configs (YAML)                 │
│                                                     │
│  default.yaml  Full feed, clickability-ranked       │
│  research.yaml arXiv + HF papers, signal-ranked     │
│  devops.yaml   Shipping tools, 7-day window         │
│  brief.yaml    PUBLISH verdict only, top 8          │
└─────────────────────────────────────────────────────┘
```

## Quick Start

```bash
# Run the ingestion pipeline
python -m oracle ingest

# Generate summaries for pending entries
python -m oracle summarize

# Render all variant editions (preview)
python -m oracle render --all --dry-run

# Render a single variant
python -m oracle render --variant research

# Deploy to production
python -m oracle render --all --webroot /var/www/html

# Check system health
python -m oracle health
```

## CLI Commands

| Command | Description |
|---------|-------------|
| `ingest` | Fetch from adapters, score, store to oracle.db |
| `summarize` | Generate summaries for entries without one |
| `query top N` | Top N entries by signal score |
| `query search "text"` | Keyword search across titles/summaries |
| `query recent --hours 24` | Recent entries |
| `render` | Render variant editions (HTML + JSON) |
| `archive` | Soft-cap archival of old entries |
| `themes` | Theme-based trend tracking |
| `dedup` | Content-hash dedup + verdict management |
| `health` | System health check |

### Ingest Options
```bash
python -m oracle ingest --sources github,arxiv    # Specific sources only
python -m oracle ingest --limit 5                 # 5 entries per source
python -m oracle ingest --dry-run                 # Fetch but don't store
```

### Render Options
```bash
python -m oracle render --list                    # Show available variants
python -m oracle render --variant brief            # Single variant
python -m oracle render --all                     # All variants
python -m oracle render --all --dry-run            # Preview mode
python -m oracle render --all --webroot /var/www   # Production deploy
```

## Scoring Engine

Pure-rule component scoring — no embeddings or LLM required.

| Component | Weight | Description |
|-----------|--------|-------------|
| Shipping | 20% | Code releases, benchmarks, working demos |
| Utility | 20% | Practical tools, frameworks, integrations |
| Replication | 25% | Reproducible research, open datasets |
| Enthusiast | 20% | Community buzz, notable figures |
| Novelty | 15% | First-of-its-kind, paradigm shifts |

**Hype Penalty:** Caps at 45% to prevent buzzwords from dominating.

**Signal Score:** 0–10 scale (per-adapter normalization → final composite).

## Source Tiers (World Monitor Pattern)

| Tier | Sources | Rationale |
|------|---------|-----------|
| **Tier 1** (PRIMARY) | arxiv, github, huggingface | Peer-reviewed research, official code releases, model registry |
| **Tier 2** (SECONDARY) | rss, hackernews | Curated tech media, curated community |
| **Tier 3** (TERTIARY) | reddit | User-generated discussion |

## Composite Verdicts

Entries are classified based on signal score + age:

| Verdict | Minimum Score | Maximum Age | Meaning |
|---------|--------------|-------------|---------|
| **PUBLISH** | ≥ 6.0 | ≤ 48h | High-signal, fresh — front page material |
| **WATCH** | ≥ 4.0 | ≤ 168h | Solid signal — worth tracking |
| **ARCHIVE** | ≥ 2.0 | ≤ 720h | Historical value — keep for reference |
| **DROP** | any | > 720h | Stale — exclude from active feeds |

## Content-Hash Dedup

SHA-256 content hashing (first 16 hex chars) for cross-source duplicate detection:
- Normalized whitespace before hashing
- Applied atomically at ingest time via `adapters/_store.py`
- Prevents the same story from appearing multiple times across sources

## Variant Editions

Each variant is a YAML config defining:

- **Filters:** verdicts, sources, min_score, max_age_h, max_items
- **Ranking:** clickability | signal_score | verdict_priority | freshness (with half-life decay)
- **Display:** theme, accent color, logo, show/hide score/tier/verdict badges

Create a new edition by adding a YAML to `variants/` — no code changes needed.

## Database

SQLite (`oracle.db`) with 784+ entries across 6 sources. Schema includes:
- Entry metadata (title, url, source, extracted_text, raw_metadata)
- Scoring (signal_score, component scores, final_score, actionability_score)
- Dedup/verdict (content_hash, verdict, source_tier)
- Summarization (summary JSON with one_liner, key_points, implications)
- Categorization (category_tags, bucket, narrative_id, topic_id)

## Sprint Log

### Sprint 0 — Foundation (2026-07-22) ✅
- Package restructure: `oracle/` + `python -m oracle` CLI
- Per-adapter timeout (10s) + threading fallback
- Source confidence tiers (3-tier system)
- Content-hash dedup (SHA-256[:16])
- Composite verdicts (PUBLISH/WATCH/ARCHIVE/DROP)
- Wired into pipeline: atomic hash/tier/verdict at insert time
- Consolidated 12 root scripts → thin wrappers + oracle/ package

### Sprint 1 — Multi-variant Editions (2026-07-22) ✅
- Variant engine: `oracle/variants.py`
- 4 default editions: default, research, devops, brief
- Variant-aware HTML/JSON renderer with theme support
- CLI: `render --variant`, `--all`, `--list`, `--webroot`
- Verified: all 4 variants render correctly with proper filtering

## Dependencies

- Python 3.10+
- PyYAML (`pip install pyyaml`)
- Trafilatura (for text extraction in adapters)
- Feedparser (RSS feeds)

## Backward Compatibility

Old scripts (`pipeline.py`, `summarize.py`, `query.py`, etc.) are thin wrappers that delegate to the unified CLI. Existing cron jobs calling `python3 pipeline.py` continue to work without changes.

## License

AGPL v3 (see LICENSE for details)