How Rechase works
end to end
A complete walkthrough of the detection pipeline — from your first API call to real-time blocks in production.
Three-layer detection pipeline
Content uploaded to your platform. Rechase intercepts via webhook or SDK before it reaches storage.
- ✓ User upload event
- ✓ Webhook call to Rechase
- ✓ < 5ms overhead
Three fingerprints computed in parallel: SHA-256 for exact matching, pHash for perceptual similarity, CLIP for semantic content.
- ✓ SHA-256 hash
- ✓ 64-bit perceptual hash
- ✓ 512-dim CLIP vector
Fingerprints compared against your database in real time. Match found → block the upload and fire an alert instantly.
- ✓ Cosine similarity check
- ✓ Hamming distance check
- ✓ Block / alert / log
Two lines to integrate
REST API or SDK. Works with any stack. No infra changes required.
Check on upload
// Before storing content const result = await rechase.check({ file: uploadedFile, platformId: "your-platform" }); if (result.matched) { return block(result.reason); }
API response
{ "matched": true, "method": "clip", "confidence": 0.94, "latency_ms": 38, "original_id": "cnt_8f2a91bc", "reason": "AI upscale of removed content" }
Three methods, one verdict
Each method catches what the others miss. Together they cover 99.3% of re-upload attempts.
First line of defence. Zero false positives. Catches lazy re-uploaders who don't bother modifying the file.
latency
< 1ms
coverage
~30%
Catches
- ✓Identical byte-for-byte re-uploads
- ✓Exact file duplicates
Misses
- ✗Any pixel change
- ✗Re-encoding or metadata strip
DCT-based 64-bit perceptual hash. Hamming distance ≤ 10 = same content. Survives most manual edits.
latency
< 5ms
coverage
~38%
Catches
- ✓Crops, rotations, resizes
- ✓Watermarks, color filters
Misses
- ✗Heavy AI generation
- ✗Style transfer, face swap
512-dim embedding from OpenAI's CLIP. Cosine similarity ≥ 0.80 = same content. Catches what every pixel-based method misses.
latency
30–50ms
coverage
~31%
Catches
- ✓AI upscale, inpainting
- ✓Style transfer, deepfakes
Misses
- ✗Completely different subject
Review flagged uploads in one place
Side-by-side comparison,
one-click decisions
Every flagged upload lands in your moderation queue with the original and the re-upload shown side by side. Fingerprint scores are surfaced inline so reviewers have full context before they act — no extra tabs, no guesswork.
- ✓All three fingerprint scores shown per item
- ✓One-click block or whitelist from the same screen
- ✓Full content ID trail linked to original removal
Queue
Original (removed)
Re-upload attempt
Every block, fully searchable
Audit trail that
never forgets
A permanent, queryable record of every blocked piece of content. Filter by method, timeframe, or user. Export for legal evidence or compliance reviews.
- ✓Filter by detection method, date range, or user ID
- ✓Confidence scores and method logged per entry
- ✓Paginated across all 1,800+ blocked items
| Content ID | Blocked | Method | Conf. | User | Actions | |
|---|---|---|---|---|---|---|
| cnt_8f2a91bc | ✗ Blocked | CLIP | 94% | usr_4421def | View → | |
| cnt_3d17fe02 | ✗ Blocked | pHash | 98% | usr_8819abc | View → | |
| cnt_aa91c344 | ✗ Blocked | CLIP | 87% | usr_2234fef | View → | |
| cnt_0011beef | ✗ Blocked | SHA-256 | 100% | usr_9910bba | View → | |
| cnt_5f82ab10 | ✗ Blocked | pHash | 96% | usr_3345ccd | View → | |
| cnt_c120de44 | ✗ Blocked | CLIP | 91% | usr_6671eef | View → |
Understand your detection patterns
Trends, method mix,
at a glance
Track block volume over time and see which detection method is doing the heavy lifting. Spot spikes before they become incidents. Tune thresholds based on real data.
- ✓7-day block trend shows spikes immediately
- ✓Method breakdown reveals attacker behaviour
- ✓False positive rate tracked automatically
Blocks per day
Method breakdown
Total blocked
1,847
This week
312
Avg confidence
93.2%
False positive rate
0.07%
Tune sensitivity to your platform
Your thresholds,
your policy
Different platforms need different sensitivity. A news aggregator tolerates more similarity than an adult content platform. Drag the sliders, toggle behaviours, save. Changes take effect on the next API call — no redeployment needed.
- ✓pHash and CLIP thresholds independently adjustable
- ✓Per-platform config — different rules for different properties
- ✓Audit log records every threshold change with timestamp
Lower = stricter. Recommended: 8–12
Higher = stricter. Recommended: 0.75–0.85
Full audit trail, real time
Blocked today
1,847
+12% vs yesterday
Detection rate
99.3%
+0.2pp this week
Avg latency
38ms
p99: 71ms
CLIP catches
31%
of all blocks
| Event ID | Time | Method | Confidence | Original | Action |
|---|---|---|---|---|---|
| EVT-4821 | 0.3s ago | CLIP | 94% | cnt_8f2a91bc | ✗ Blocked |
| EVT-4820 | 12s ago | pHash | 98% | cnt_3d17fe02 | ✗ Blocked |
| EVT-4819 | 1m ago | CLIP | 87% | cnt_aa91c344 | ✗ Blocked |
| EVT-4818 | 4m ago | SHA-256 | 100% | cnt_0011beef | ✗ Blocked |
| EVT-4817 | 9m ago | pHash | 96% | cnt_5f82ab10 | ✗ Blocked |
99.3%
Detection rate
38ms
Avg response time
< 1 day
Typical integration
Your team knows instantly
Slack alerts with full context — method, confidence, link to original. No dashboard login required.
Or receive raw webhook
{ "event": "content.blocked", "matched": true, "method": "clip", "confidence": 0.94, "upload_id": "upl_7f3b91", "original_id": "cnt_8f2a91bc", "user_id": "usr_4421def", "timestamp": "2026-04-20T14:47:03Z" }