Ops Console
FishID Ops
Machines
↻ Refresh all
CPU · Always On · serves /api/segment-only
—
loading…
Specs
Pipeline
Tunnel
GPU · Experiments only · not called by any production endpoint
—
loading…
Specs
Pipeline
Tunnel
Architecture
/api/identify?engine=v1|v2|v3 is a single edge entry that dispatches to one of three pure-VLM engines
— none of them call DINO/SAM. Cutout is opt-in via ?segment=true,
which fans out to /api/segment in parallel and proxies the CPU server's
/api/segment-only. The GPU instance is no longer in any production code path
(kept available for ad-hoc experiments only).
Segment Pipeline
Used only by
/api/segment standalone calls · identify engines do NOT use this
GroundingDINO · Detection Model
—
loading…
SAM · CPU Segmentation Model
—
loading…
Engine
V1 → KV
vlm_prompts · V2 → engine_config_v2 · V3 → engine_config_v3
V1 Prompt
Read-only display of active version · KV
vlm_prompts
Active version
—
System Message
User Prompt
Max Tokens
V2 Prompt
Read-only display of active version · KV
engine_config_v2
Active version
—
System Prompt
User Prompt
Max Tokens
V3 Tuning
Stage B candidate recall + low-confidence threshold
Max Candidates
cap on candidates passed from Stage B (recall) to Stage C
Low Confidence Threshold
flags
low_confidence:true on result when Stage C confidence < this
V3 Stage A · Visual Feature Extraction
Read-only · VLM extracts features → code recalls candidates by family / open_guess / habitat
Active version
—
System Prompt
User Prompt
Max Tokens
V3 Stage C · Pick from Candidates
Read-only · VLM picks one species_id from the recalled candidate list (or unknown_species)
Active version
—
System Prompt
User Prompt (candidates JSON appended at runtime)
Max Tokens
Identifications
All V1/V2/V3 calls log to D1 · model prefix marks version
Stage A 准确率（命中样本中，对照真实物种）
Loading…
Events (last 7 days)
Billing
Account Balance
—
—
Last 7 Days Spend
—
CPU ¥0.44/h (c4il.xlarge) · GPU billed only when running