face-intel / docs /ARCHITECTURE.md

Marwan

Restructure + add reverse face search (PimEyes-style)

f5eeb1c 21 days ago

25.2 kB

Face Intel — Architecture

Refactored architecture (v2). Introduces a services layer, a pipeline package, a provider registry, a metrics subsystem, separated storage, shared domain models, strict one-way dependency direction, comprehensive dependency injection, and structured execution-context logging.

1. Dependency Graph (strict one-way)

                        ┌──────────┐
                        │   API    │   FastAPI routes + middleware
                        └────┬─────┘
                             │
                        ┌────▼─────┐
                        │ Services │   business logic (8 services)
                        └────┬─────┘
                             │
              ┌──────────────┼──────────────┐
              │              │              │
         ┌────▼─────┐  ┌────▼─────┐  ┌─────▼──────┐
         │ Orchest. │  │ Confid.  │  │ Normaliz. │
         └────┬─────┘  └────┬─────┘  └─────┬──────┘
              │             │              │
              └─────────────┼──────────────┘
                            │
                       ┌────▼─────┐
                       │ Pipeline │   validation → preprocess → hash → feature_extract
                       └────┬─────┘
                            │
                       ┌────▼─────┐
                       │ Providers│   detection / recognition / scraper / reverse
                       └────┬─────┘
                            │
                       ┌────▼─────┐
                       │ Storage  │   database / cache / artifacts / reference_store
                       └────┬─────┘
                            │
                       ┌────▼─────┐
                       │  Utils   │   image / http / audit / timing / logging
                       └────┬─────┘
                            │
                       ┌────▼─────┐
                       │  Models  │   shared domain DTOs (pure pydantic, no deps)
                       └──────────┘

Rule: arrows point downward. A layer may import only from layers below it. models/ is at the bottom and depends on nothing but pydantic

stdlib. No circular imports — verified by scripts/check_imports.py.

Cross-cutting: config/ (settings) is read by every layer but is a value object, not stateful. metrics/ is consumed by orchestrator + services + API but owns no upstream dependencies.

2. Package Responsibilities

Package	Responsibility	Key Exports
`config/`	Environment-driven settings (Pydantic BaseSettings)	`Settings`, `make_settings()`
`models/`	Shared domain DTOs (pure pydantic, no logic)	`Job`, `JobRequest`, `UnifiedFaceReport`, `ProviderInfo`, `HealthSnapshot`, `APIResponse`
`utils/`	Stateless helpers: image I/O, HTTP session, audit log, timing, structured logging	`bytes_to_numpy`, `shared_session`, `audit_log`, `execution_context`, `setup_logging`
`storage/`	Persistence: SQLite (jobs), TTL cache, file artifacts, reference gallery	`Database`, `Cache`, `ArtifactStore`, `ReferenceStore`
`metrics/`	Per-provider latency/success/retry counters, timings, health, cache hit ratio	`MetricsCollector` (facade)
`providers/`	Provider Protocol + BaseProvider + Registry + 4 capability folders	`Provider`, `BaseProvider`, `ProviderRegistry`, `PROVIDER_MANIFEST`
`pipeline/`	Local pre-orchestrator stages: validate → preprocess → hash → feature-extract → postprocess	`InputValidator`, `ImagePreprocessor`, `ImageHasher`, `FeatureExtractor`, `PipelineOutput`
`orchestrator/`	Async fan-out, retry, circuit breaker	`Orchestrator`, `RetryPolicy`, `HealthMonitor`
`normalization/`	Merge multi-provider results into unified report; internal DTOs only	`ReportMerger`, `NormalizedBox`, `NormalizedMatch`
`confidence/`	Explainable scoring + cross-provider conflict detection	`ConfidenceEngine`, `ConflictDetector`, `Explainer`
`services/`	Business logic; one service per concern	`DetectionService`, `RecognitionService`, `SearchService`, `JobService`, `ProviderService`, `CacheService`, `HealthService`, `ExportService`
`api/`	FastAPI app, DI container, middleware, route handlers	`create_app()`, `ServiceContainer`, `build_container()`
`ui/`	Static SPA (served by FastAPI)	—
`tests/`	Unit + provider + integration test suites	—

3. Execution Flow

HTTP request
    │
    ▼
FastAPI middleware:  request-id → rate-limit → CORS
    │
    ▼
Route handler  (api/routes/<resource>.py)
    │  depends( get_<service> )  →  pulls service off app.state.container
    ▼
Service  (services/<name>_service.py)
    │  1. InputValidator.validate(image_url|base64|bytes)
    │  2. ImagePreprocessor.from_url|from_bytes  →  PreprocessedImage
    │  3. ImageHasher.hash                        →  cache key
    │  4. FeatureExtractor.extract                →  PipelineOutput
    │                                              (image + hash + face_crops + optional gallery/scrape_url)
    │  5. Orchestrator.run(pipeline_output, capabilities)
    │  6. ReportMerger.merge(results)             →  UnifiedFaceReport
    │  7. return {report, elapsed_ms}
    ▼
Orchestrator  (orchestrator/runner.py)
    │  for each provider matching capabilities:
    │    - skip if circuit breaker open
    │    - check cache; hit → return cached ProviderResult
    │    - else: asyncio.to_thread(provider.execute(pipeline_output))
    │    - apply RetryPolicy (exponential backoff + jitter)
    │    - record metrics (latency, success/failure, retries)
    │    - cache successful results
    ▼
Provider  (providers/<category>/<name>.py)
    │  receives PipelineOutput
    │  extracts .image / .face_crops / .gallery / .scrape_url
    │  runs its model / HTTP call
    │  returns ProviderResult(raw=verbatim, normalized={...}, elapsed_ms, success)
    ▼
Normalization  (normalization/merger.py)
    │  collects NormalizedBox / NormalizedMatch / NormalizedScrapeImage / NormalizedReverseMatch
    │  preserves every ProviderResult as Evidence
    │  delegates scoring to ConfidenceEngine
    │  delegates conflicts to ConflictDetector
    │  returns UnifiedFaceReport
    ▼
Response  →  JSON to client

4. Lifecycle Diagrams

4.1 Application Startup

create_app(settings)
    │
    ├─ setup_logging(settings)               # loguru sinks + execution-context format
    │
    ├─ lifespan:
    │     build_container(settings)          # composition root (DI)
    │       │
    │       ├─ ProviderRegistry(settings)
    │       │     .discover()                 # imports each provider module, instantiates, registers
    │       │                                 # missing optional deps → NOT_CONFIGURED (graceful)
    │       │
    │       ├─ Cache(ttl, max_entries)
    │       ├─ Database(path)                 # SQLite, schema migrated
    │       ├─ ArtifactStore(root=uploads/)
    │       ├─ ReferenceStore(root=gallery/)
    │       │
    │       ├─ MetricsCollector(failure_threshold, recovery_seconds)
    │       ├─ HealthMonitor(metrics.health)
    │       │
    │       ├─ Orchestrator(registry, cache, metrics, health, settings, retry_policy)
    │       │
    │       ├─ ConfidenceEngine()
    │       ├─ ConflictDetector()
    │       │
    │       ├─ InputValidator()
    │       ├─ ImagePreprocessor(max_dim=1024)
    │       ├─ ImageHasher()
    │       ├─ FeatureExtractor(detector=first_detection_provider)
    │       │
    │       ├─ DetectionService(...)
    │       ├─ RecognitionService(...)
    │       ├─ SearchService(...)
    │       ├─ JobService(detection, recognition, search, database, metrics)
    │       ├─ ProviderService(registry)
    │       ├─ CacheService(cache)
    │       ├─ HealthService(registry, health, metrics)
    │       └─ ExportService(database)
    │
    ├─ app.state.container = container
    │
    └─ mount routes + UI

4.2 Request Lifecycle

1.  HTTP request hits FastAPI
2.  RequestContextMiddleware  →  assigns X-Request-ID, starts timer
3.  RateLimitMiddleware       →  enforces per-IP limit (sliding 60s window)
4.  Route handler             →  Depends(get_<service>) pulls from container
5.  Service                   →  runs pipeline → orchestrator → normalization → confidence
6.  Orchestrator              →  per-provider: cache check → retry-wrapped invoke → metrics
7.  Provider                  →  executes within execution_context(provider_id=...)
8.  Result flows back up:     ProviderResult → UnifiedFaceReport → JSON response
9.  Response headers          →  X-Request-ID, X-Response-Time-ms

4.3 Circuit Breaker Lifecycle

provider.invoke() succeeds
    │
    ▼
HealthMetrics.record_success(name)
    │  → consecutive_failures = 0
    │  → avg_latency_ms = exp-avg
    │  → if circuit_open: close it
    ▼
circuit stays CLOSED


provider.invoke() fails
    │
    ▼
HealthMetrics.record_failure(name)
    │  → consecutive_failures += 1
    │  → if consecutive_failures >= threshold AND not circuit_open:
    │        circuit_open = True
    │        circuit_opened_at = now()
    ▼
circuit OPEN  →  orchestrator skips this provider
                   until recovery_seconds elapsed


next invoke attempt after recovery_seconds
    │
    ▼
HealthMetrics.is_circuit_open(name)
    │  → if circuit_open AND now - opened_at > recovery_seconds:
    │        circuit_open = False   (half-open)
    │        return False           (allow attempt)
    ▼
provider gets one trial invoke
    success → circuit stays closed
    failure → circuit re-opens

5. Extension Guide — Adding a New Provider

Adding a provider requires two changes and zero orchestrator/API edits.

Step 1 — Implement the provider

Create providers/<category>/<your_provider>.py:

from __future__ import annotations
from config.settings import Settings, settings as _default
from pipeline.feature_extraction import PipelineOutput
from providers.base import BaseProvider, ProviderCapability, ProviderResult


class YourProvider(BaseProvider):
    name = "your_provider"
    capability = ProviderCapability.DETECTION  # or RECOGNITION / SCRAPING / REVERSE_SEARCH

    def __init__(self, settings: Settings | None = None) -> None:
        super().__init__(settings=settings or _default)
        # ... initialize your model / client

    def is_available(self) -> bool:
        # Return False if optional deps or API keys are missing.
        return True

    def _run(self, pipeline_output: PipelineOutput) -> tuple[dict, dict]:
        img = pipeline_output.image   # extract what you need
        # ... do your work
        raw = {"...": ...}            # verbatim response (preserved as evidence)
        normalized = {
            "boxes": [{"x": 0, "y": 0, "w": 0, "h": 0}],
            "num_faces": 1,
            "confidences": [0.95],
            "landmarks": None,
        }
        return raw, normalized

Step 2 — Add one manifest entry

Edit providers/registry.py:

PROVIDER_MANIFEST: list[ManifestEntry] = [
    # ... existing entries ...
    ManifestEntry(
        name="your_provider",
        module_path="providers.detection.your_provider",
        class_name="YourProvider",
        capability=ProviderCapability.DETECTION,
        enable_flag="enable_your_provider",
        description="Your provider — one-line description",
        optional_dependency=True,   # set True if it imports an optional package
    ),
]

Step 3 — Add the enable flag

Edit config/settings.py:

enable_your_provider: bool = False
# ... any tuning knobs your provider needs, e.g.:
your_provider_threshold: float = 0.7

Step 4 — Done

The registry auto-discovers your provider on startup.
The orchestrator invokes it for matching capability queries.
/providers lists it automatically.
/stats tracks its latency, success rate, retries.
The circuit breaker protects against its failures.
The cache dedupes identical invocations.

No other file needs to change. This is the core extensibility guarantee.

6. Dependency Injection

Principle

No global objects for stateful services. Inject services, storage, metrics, cache, and orchestrator through constructors.

Composition Root

api/container.py::build_container() is the only place where services are constructed. It runs once at app startup and stores the ServiceContainer on app.state.container.

How Routes Receive Services

# api/routes/faces.py
from fastapi import Depends
from api.deps import get_detection_service

@router.post("/detect")
async def detect_faces(req: FaceRequest,
                       svc: DetectionService = Depends(get_detection_service)):
    return await svc.detect(job_req)

get_detection_service is a thin wrapper:

# api/deps.py
def get_detection_service(request: Request):
    return get_container(request).detection_service

Why This Matters

Testability: tests construct a container with :memory: DB, disabled network providers, and assert against the same code path as production.
No hidden state: every dependency is visible in the constructor signature — no surprise module-level singletons.
Easy overrides: swap any component (cache, DB, registry) by constructing a custom container in tests or for feature flags.

7. Structured Logging

Every log line produced inside an execution carries:

2026-07-10 14:30:00.123 | INFO     | eid=abc123def456 | pid=haar | retry=0 | status=started | invoking haar

Field	Source	Meaning
`eid`	`execution_context(execution_id=...)`	Unique id for the top-level job
`pid`	`execution_context(provider_id=...)`	Provider name (or `-` for service-level logs)
`retry`	`execution_context(retry_count=...)`	Retry attempt number (0 = first try)
`status`	updated via the context dict	`started` → `running` → `success` / `failed` / `retried`

Usage

from utils.logging import execution_context, new_execution_id

async def my_handler():
    eid = new_execution_id()
    with execution_context(execution_id=eid, provider_id="my_service"):
        logger.info("started")           # → eid=…, pid=my_service, status=started
        # ... do work ...
        logger.info("completed")          # → same eid + pid

JSON Mode

Set FI_LOG_JSON=true to emit newline-delimited JSON for log aggregators (Loki, Datadog, CloudWatch):

{"timestamp":"2026-07-10T14:30:00.123Z","level":"INFO","execution_id":"abc123def456","provider_id":"haar","retry_count":0,"status":"started","message":"invoking haar"}

8. Metrics Subsystem

metrics/ tracks six categories consumed by /stats and the UI:

Subsystem	What it tracks	Where it's recorded
`ProviderMetrics`	per-provider invocations, successes, failures, retries, latency samples	orchestrator after each provider call
`TimingCollector`	per-operation duration histograms (p50, p95)	services record `job.detection`, `job.recognition`, `job.search`
`CounterRegistry`	global counters (cache.hits, cache.misses, retries., failures., jobs..completed)	orchestrator + services
`HealthMetrics`	per-provider consecutive failures, avg latency, circuit-breaker state	orchestrator via HealthMonitor

MetricsCollector is a facade that owns all four — injected as a single object so services don't depend on four separate classes.

GET /stats returns the full snapshot:

{
  "providers": [{"name": "haar", "invocations": 42, "successes": 41, "failures": 1, "retries": 0, "avg_latency_ms": 6.2, "p95_latency_ms": 12.4, "success_rate": 0.976}],
  "timings": {"job.detection": {"count": 42, "avg_ms": 14.3, "p50_ms": 12.0, "p95_ms": 28.1}},
  "counters": {"cache.hits": 18, "cache.misses": 24, "jobs.detection.completed": 42},
  "health": [{"name": "haar", "consecutive_failures": 0, "avg_latency_ms": 6.2, "circuit_open": false}]
}

9. REST API Surface

Method	Path	Description
`GET`	`/health`	Liveness probe
`GET`	`/health/live`	Liveness
`GET`	`/health/ready`	Readiness
`GET`	`/health/providers`	Per-provider health snapshot
`GET`	`/stats`	Full metrics snapshot
`GET`	`/providers`	List all providers + manifest errors
`GET`	`/providers/{name}`	Single provider info
`GET`	`/cache`	Cache stats (entries, hit ratio, evictions)
`DELETE`	`/cache`	Clear cache
`POST`	`/jobs`	Create + run a job (detection / recognition / search / full_pipeline)
`GET`	`/jobs`	List recent jobs
`GET`	`/jobs/{id}`	Job metadata
`GET`	`/jobs/{id}/result`	Job result (UnifiedFaceReport)
`POST`	`/faces/detect`	Convenience: detect-only
`POST`	`/faces/recognize`	Convenience: recognize-only
`GET`	`/faces/gallery`	List known persons
`DELETE`	`/faces/gallery/{name}`	Remove a known person
`POST`	`/search/reverse`	Reverse image search
`POST`	`/search/scrape`	Scrape images from a URL
`GET`	`/export/{job_id}`	Download job result as JSON
`GET`	`/docs`	OpenAPI Swagger UI

10. File Tree (refactored)

face-intel/
├── app.py                          # entry point: uvicorn app:app
├── requirements.txt
├── .env.example
├── README.md
│
├── config/
│   ├── __init__.py
│   └── settings.py                 # Settings + make_settings() factory
│
├── models/                         # NEW: shared domain DTOs
│   ├── __init__.py
│   ├── jobs.py                     # Job, JobRequest, JobStatus, JobResult, JobKind
│   ├── providers.py                # ProviderCapability, ProviderStatus, ProviderInfo, ProviderConfig
│   ├── reports.py                  # UnifiedFaceReport, FaceDetection, FaceMatch, Evidence, ConfidenceScore, ConflictReport
│   ├── responses.py                # APIResponse, PaginatedResponse, ErrorResponse, HealthResponse
│   └── health.py                   # HealthSnapshot, ProviderHealthSnapshot, SystemHealthSnapshot
│
├── utils/                          # lowest layer
│   ├── __init__.py
│   ├── image.py                    # bytes↔numpy, BBox, crop, hash, draw
│   ├── http.py                     # shared requests.Session with retries
│   ├── audit.py                    # append-only JSONL audit log
│   ├── timing.py                   # @contextmanager timed()
│   └── logging.py                  # NEW: loguru + execution_context()
│
├── storage/                        # separated responsibilities
│   ├── __init__.py
│   ├── database.py                 # SQLite (jobs, results)
│   ├── cache.py                    # in-memory TTL + LRU
│   ├── artifacts.py                # filesystem: uploads/, generated/
│   └── reference_store.py          # known-faces gallery (.npy + manifest.json)
│
├── metrics/                        # NEW
│   ├── __init__.py
│   ├── provider_metrics.py         # per-provider latency/success/retry
│   ├── timings.py                  # operation-level p50/p95
│   ├── counters.py                 # named integer counters
│   ├── health_metrics.py           # circuit breaker state
│   └── collector.py                # MetricsCollector facade
│
├── providers/
│   ├── __init__.py                 # re-exports (no globals)
│   ├── base.py                     # Provider Protocol, BaseProvider, ProviderResult
│   ├── registry.py                 # NEW: ProviderRegistry class + PROVIDER_MANIFEST
│   ├── detection/
│   │   ├── __init__.py
│   │   └── haar.py                 # implemented (DI-updated)
│   ├── recognition/                # stubs (to be implemented Phase 3)
│   ├── scraper/
│   └── reverse/
│
├── pipeline/                       # NEW
│   ├── __init__.py
│   ├── validation.py               # InputValidator
│   ├── preprocessing.py            # ImagePreprocessor → PreprocessedImage
│   ├── hashing.py                  # ImageHasher (SHA-256 cache key)
│   ├── feature_extraction.py       # FeatureExtractor → PipelineOutput
│   └── postprocessing.py           # dedupe, clamp, filter
│
├── orchestrator/
│   ├── __init__.py
│   ├── runner.py                   # Orchestrator (async fan-out, DI-injected)
│   ├── retry.py                    # RetryPolicy, with_retry_sync/async
│   └── health.py                   # HealthMonitor (circuit breaker gate)
│
├── normalization/
│   ├── __init__.py
│   ├── schema.py                   # internal DTOs (NormalizedBox, NormalizedMatch, ...)
│   └── merger.py                   # ReportMerger (provider results → UnifiedFaceReport)
│
├── confidence/
│   ├── __init__.py
│   ├── engine.py                   # ConfidenceEngine (weighted sub-scores)
│   ├── explainer.py                # human-readable explanations
│   └── conflicts.py                # ConflictDetector (cross-provider disagreements)
│
├── services/                       # NEW
│   ├── __init__.py
│   ├── detection_service.py
│   ├── recognition_service.py
│   ├── search_service.py
│   ├── job_service.py              # full-pipeline + persistence
│   ├── provider_service.py
│   ├── cache_service.py
│   ├── health_service.py
│   └── export_service.py
│
├── api/
│   ├── __init__.py
│   ├── main.py                     # create_app() factory
│   ├── container.py                # build_container() — composition root
│   ├── deps.py                     # FastAPI Depends() callables
│   ├── middleware.py               # request-id, rate limit
│   └── routes/
│       ├── __init__.py
│       ├── health.py
│       ├── stats.py
│       ├── providers.py
│       ├── cache.py
│       ├── jobs.py
│       ├── faces.py
│       ├── search.py
│       └── export.py
│
├── ui/
│   └── static/                     # served by FastAPI (to be built Phase 8)
│
├── tests/                          # to be built Phase 9
│   ├── unit/
│   ├── providers/
│   └── integration/
│
└── docs/
    └── ARCHITECTURE.md             # this file

11. Verification

The refactor has been smoke-tested end-to-end:

✓ All 11 layers import cleanly — no circular dependencies
✓ DI container builds with all 15 manifest entries registered
✓ FastAPI app boots with 32 routes
✓ End-to-end detection job runs: pipeline → orchestrator → haar → normalization → confidence → report
✓ Haar provider correctly receives PipelineOutput (not raw numpy)
✓ Evidence preserved with raw + normalized + elapsed_ms + success
✓ Confidence engine produces explainable sub-scores
✓ Metrics subsystem records invocations, successes, timings
✓ Circuit-breaker state queryable via /health/providers

12. What's Next

The refactor establishes the production-grade skeleton. The remaining work from the original 10-phase plan:

Phase 3 (continued): implement the 14 remaining provider modules (dnn, mtcnn, retinaface, face_recognition, deepface, insightface, beautifulsoup, selenium, bing, duckduckgo, google_lens, serpapi, yandex, tineye). Each follows the haar.py pattern — ~50 LOC.
Phase 8: build the UI SPA in ui/static/.
Phase 9: write the test suites in tests/.
Phase 10: provider-specific docs, deployment, benchmarks.

The architecture will not need to change for any of these — only additions within existing packages.