CAPI AI — AOI PatchCore Inspection Platform

Production-oriented AI inference service for AOI panel inspection. It receives AOI requests over TCP, runs the configured PatchCore pipeline, returns the legacy AOI result together with the QJPG report, and stores traceable results for Web review.

🇹🇼 繁體中文說明 → README.zh-TW.md

What this repository contains

• Inference server — capi_server.py handles persistent TCP client connections, request parsing, model dispatch, inference, and protocol responses.
• PatchCore pipeline — capi_inference.py and capi_preprocess.py cover panel preprocessing, tile/zone routing, anomaly scoring, heatmaps, MARK and bomb handling, and post-processing rules defined by model configuration.
• Traceability and Web UI — capi_database.py stores inference, image, and tile records in SQLite; capi_web.py serves monitoring, search, record details, RIC review, and administration pages.
• Training and model library — the /training → /train/new workflow prepares training data, reviews tiles, trains model bundles, and manages activation from /models.
• Deployment support — release metadata, deploy ZIP generation, manual update, and pull-based update helpers are included in the repository.

The production data path is:

AOI client
    │ TCP
    ▼
capi_server.py ──► capi_inference.py / capi_preprocess.py
    │                              │
    ├── legacy AOI + QJPG response │
    ├── SQLite inference records   └── heatmaps and diagnostics
    ▼
capi_web.py ──► dashboard, review, training, model library, settings

Requirements and installation

Use Python 3.10 or newer; the current development/deployment environments use Python 3.11/3.12.

python -m pip install -r requirements.txt

The repository does not contain production model weights. A runnable installation also needs model bundles and image-path mappings appropriate for the target machine. Weight files, databases, local datasets, and local credentials are intentionally excluded from normal source control and deployment packaging.

Start the server

Windows local test

server_config_local.yaml is the local profile:

• TCP server: 0.0.0.0:7891
• Web UI: http://localhost:8080
• SQLite database: ./test_results.db
• Heatmaps: ./test_heatmaps

Start it with either command:

python capi_server.py --config server_config_local.yaml
# or
start_server_local.bat

If a panel dataset is available, auto_sender.py can send sample requests:

python auto_sender.py --host 127.0.0.1 --port 7891 --ng-folder D:\path\to\panels --count 1

Linux production

Edit server_config.yaml for the target machine, then use the service helper:

chmod +x start_server.sh
./start_server.sh              # stop old process, start in background, tail the log
./start_server.sh status
./start_server.sh log
./start_server.sh stop

The production profile currently defaults to TCP port 7907 and Web port 80. The actual ports, database path, heatmap path, model list, path mapping, retention policy, and optional integrations are controlled by server_config.yaml.

For a direct foreground start:

python3 capi_server.py --config server_config.yaml

Do not copy production paths or credentials into the local profile. In particular, server_config.yaml contains machine-specific paths and MES settings that must be reviewed before deployment.

TCP protocol

The server accepts semicolon-delimited AOI@ requests. A request without bomb coordinates is:

AOI@<glass_id>;<model_id>;<machine_no>;<resolution_x>,<resolution_y>;<machine_judgment>;<image_dir>

A request with bomb data adds an image prefix and coordinates before the image path:

AOI@<glass_id>;<model_id>;<machine_no>;<resolution_x>,<resolution_y>;<machine_judgment>;<image_prefix>;<coordinates>;<image_dir>

machine_judgment is normally OK, NG, or HY. HY skips AI inference and is returned as an image-abnormal result.

The current response is CRLF-terminated and contains both formats, in this order:

AOI@<glass_id>;<model_id>;<machine_no>;<machine_judgment>;<ai_judgment>
@QJPG-<glass_id>;<mark_status>;<mark_text>;<defect_field>,

Clients should identify each line by its prefix (AOI@ or @QJPG-) instead of assuming that a response contains only one line. ai_judgment can be OK, NG, or ERR:<description>; the internal OK-i result is exposed as OK in the legacy response. The complete field and QJPG defect-code specification is in docs/client_communication_protocol.zh-TW.md.

Main Web UI entry points

Open http://<server>:<web_port>/ after the server starts.

Path	Purpose
/	Live dashboard and current shift status
/search	Search and export inference records
/record/<id>	Record details, images, tiles, and heatmaps
/ric	RIC, over-review, miss-review, MES comparison, and related reports
/ric/within-spec-logs	Within-spec review list and details
/training	Training hub
/train/new	New-machine PatchCore training workflow
/models	Model bundle inspection and activation
/debug	Single-image and coordinate diagnostics
/white-frame	White-frame overview and records
/settings	Authenticated settings and account administration
/logs	Server log viewer
/release-notes	In-app release notes
/api/status	Runtime and hardware status JSON
/api/version	Deployed version and build metadata JSON

Configuration boundaries

File or directory	Responsibility
server_config.yaml	Production TCP/Web settings, SQLite, heatmaps, path mapping, model list, cleanup, training, and optional integrations
server_config_local.yaml	Windows/local profile with local ports and output paths
configs/capi_3f.yaml	Legacy/fallback model configuration, image-prefix mappings, thresholds, exclusion zones, bomb rules, and post-processing
model/<machine>-<timestamp>/	Bundles produced by the training workflow; each bundle contains its own model configuration and metadata
VERSION / CHANGELOG.md	Release identity and operator-facing change history

Production model_configs should point to the bundle machine_config.yaml files that match the incoming ModelID. configs/capi_3f.yaml is retained for legacy/fallback use; it is not a substitute for installing the required model weights.

Common development checks

Run the protocol smoke test without starting a listener:

python -X utf8 capi_server.py --test-protocol

Run the automated test suite from the repository root:

python -m pytest tests/

Related documentation

• Client communication protocol
• New-machine model training SOP
• PatchCore training architecture
• Experimental pull-based update workflow
• Deployment ZIP builder
• Central dashboard
• Change history

Internal project; not intended for public distribution.