# eBay MPN Matcher

Multi-step pipeline: eBay search → exact MPN match → AI match → price comparison.

## Setup

```bash
npm install
cp .env.example .env
# Fill in EBAY_CLIENT_ID, EBAY_CLIENT_SECRET, OPENAI_API_KEY
```

## Files required

| File | Description |
|------|-------------|
| `keywords.txt` | One search keyword per line |
| `prices.xlsx` | Col A: partnumber, Col B: price (USD) |
| `system_prompt.txt` | AI system prompt for this product profile |

## Run

```bash
node index.js
```

Output saved to `./output/`:
- `results_<runId>.xlsx` — full results with PASS/FAIL
- `results_<runId>.json` — same data as JSON
- `run_<runId>.log`      — full run log with timing
- `ai_cache.json`        — cached AI matches (reused on next run)

## Flow

```
keywords.txt
    │
    ▼ Step 1: eBay search (all keywords, dedup by URL item ID)
    │
    ▼ Step 2: Exact MPN match (regex + normalized)
    │         ├─ matched   → list A
    │         └─ unmatched → Step 3
    │
    ▼ Step 3: AI match (gpt-4o-mini, batch 40 items)
    │         ├─ GOOD_MATCH / VARIANT_MISMATCH → list B
    │         ├─ INSUFFICIENT_DATA → fetch detail → retry AI
    │         └─ UNRELATED_PRODUCT → skip
    │
    ▼ Step 4: Merge A+B → compare price → PASS/FAIL → export
```

## Match methods in output

| Method | Description |
|--------|-------------|
| `exact` | Regex found MPN in title/specs |
| `ai_exact` | AI matched with GOOD_MATCH |
| `ai_variant` | AI matched with VARIANT_MISMATCH |
| `ai_detail` | AI matched after fetching detail page |
| `ai_cached` | From previous run cache |

## Config (.env)

| Key | Default | Description |
|-----|---------|-------------|
| `AI_BATCH_SIZE` | 40 | Listings per AI request |
| `AI_CONFIDENCE_THRESHOLD` | 50 | Min confidence to count as PASS |
| `OPENAI_MODEL` | gpt-4o-mini | OpenAI model |
| `EBAY_MARKETPLACE` | EBAY_US | eBay marketplace ID |

## Adding a new product profile

1. Create new `keywords_<profile>.txt`
2. Create new `system_prompt_<profile>.txt` explaining the product domain
3. Update `.env` to point to new files
4. Run — AI cache is per-itemId so it auto-separates