fix(inference): METADATA-35 fit Call 1/2 images to 1024 long edge
Downscale processed PNGs only on the inference send path (keep alpha, never upscale). Stored inventory/Woo blobs unchanged. Adds Pillow + unit/e2e-style tests; cite tests on METADATA-35.
This commit is contained in:
+45
-11
@@ -5,6 +5,7 @@ Cites METADATA-35..115, SETTINGS-30/45/90.
|
||||
|
||||
import base64
|
||||
import hashlib
|
||||
import io
|
||||
import json
|
||||
import time
|
||||
import urllib.error
|
||||
@@ -13,12 +14,53 @@ from typing import Any
|
||||
|
||||
from fastapi import APIRouter
|
||||
from fastapi.responses import JSONResponse
|
||||
from PIL import Image
|
||||
|
||||
from .db import get_connection
|
||||
from .crypto import decrypt
|
||||
|
||||
router = APIRouter(prefix="/api/v1")
|
||||
|
||||
# METADATA-35: Call 1 (and Call 2) send clean processed photos at 1024 px long edge or less.
|
||||
_INFER_MAX_LONG_EDGE = 1024
|
||||
|
||||
|
||||
def fit_long_edge_png(png_bytes: bytes, max_long_edge: int = _INFER_MAX_LONG_EDGE) -> bytes:
|
||||
"""Downscale a PNG so max(width, height) <= max_long_edge. Never upscale. Keep alpha.
|
||||
|
||||
Used only for the inference request payload; stored processed_blob / Woo media stay full size.
|
||||
Unreadable or non-image bytes are returned unchanged so a bad blob cannot crash Call 1.
|
||||
"""
|
||||
if not png_bytes or max_long_edge < 1:
|
||||
return png_bytes
|
||||
try:
|
||||
im = Image.open(io.BytesIO(png_bytes))
|
||||
im.load()
|
||||
except Exception:
|
||||
return png_bytes
|
||||
w, h = im.size
|
||||
long_edge = max(w, h)
|
||||
if long_edge <= max_long_edge:
|
||||
# Already within budget; re-encode only if we need a stable PNG (caller always base64s).
|
||||
return png_bytes
|
||||
scale = max_long_edge / float(long_edge)
|
||||
nw = max(1, int(round(w * scale)))
|
||||
nh = max(1, int(round(h * scale)))
|
||||
# Preserve alpha: convert palette/etc. to RGBA when needed, LANCZOS downscale.
|
||||
if im.mode not in ("RGB", "RGBA", "L", "LA"):
|
||||
im = im.convert("RGBA")
|
||||
out = im.resize((nw, nh), Image.Resampling.LANCZOS)
|
||||
buf = io.BytesIO()
|
||||
out.save(buf, format="PNG", optimize=True)
|
||||
return buf.getvalue()
|
||||
|
||||
|
||||
def _image_part_from_processed(png_bytes: bytes) -> dict:
|
||||
"""Build one OpenAI-style image_url part from a processed PNG (METADATA-35 sized)."""
|
||||
fitted = fit_long_edge_png(png_bytes)
|
||||
b64 = base64.b64encode(fitted).decode()
|
||||
return {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{b64}"}}
|
||||
|
||||
CALL1_SEED = (
|
||||
"Classify the slab's species, wood categories, edge type, and figure from the supplied "
|
||||
"photos, choosing ONLY from the supplied synced taxonomy lists (invent no category, term, "
|
||||
@@ -203,17 +245,12 @@ async def call1(slab_id: str) -> dict:
|
||||
client = InferenceClient(base, key, model)
|
||||
act = _active_prompt(conn, "call1")
|
||||
prompt_text = act[1] if act else CALL1_SEED
|
||||
# Build vision message with processed photos (1024px) + taxonomy lists.
|
||||
# Build vision message with processed photos (METADATA-35: long edge <= 1024) + taxonomy.
|
||||
photos = conn.execute(
|
||||
"SELECT processed_blob FROM slab_photos WHERE slab_id=? AND processed_blob IS NOT NULL ORDER BY seq LIMIT 8",
|
||||
(slab_id,),
|
||||
).fetchall()
|
||||
image_parts = []
|
||||
for p in photos:
|
||||
image_parts.append({
|
||||
"type": "image_url",
|
||||
"image_url": {"url": f"data:image/png;base64,{base64.b64encode(p[0]).decode()}"},
|
||||
})
|
||||
image_parts = [_image_part_from_processed(p[0]) for p in photos]
|
||||
tax = _taxonomy_lists(conn)
|
||||
text = (f"{prompt_text}\n\nSYNCED TAXONOMY (only legal choices):\n{json.dumps(tax)}\n\n"
|
||||
f"SKU={row['sku']} length_in={row['length_in']} thickness_in={row['thickness_in']}")
|
||||
@@ -359,10 +396,7 @@ async def call2(slab_id: str) -> dict:
|
||||
"SELECT processed_blob FROM slab_photos WHERE slab_id=? AND processed_blob IS NOT NULL ORDER BY seq LIMIT 8",
|
||||
(slab_id,),
|
||||
).fetchall()
|
||||
image_parts = [
|
||||
{"type": "image_url", "image_url": {"url": f"data:image/png;base64,{base64.b64encode(p[0]).decode()}"}}
|
||||
for p in photos
|
||||
]
|
||||
image_parts = [_image_part_from_processed(p[0]) for p in photos]
|
||||
brand = conn.execute("SELECT value FROM settings WHERE key='brand_voice'").fetchone()
|
||||
geo = conn.execute("SELECT value FROM settings WHERE key='geo_context'").fetchone()
|
||||
text = (f"{prompt_text}\n\nDIMENSION FACT BLOCK (exact, invent no numbers):\n"
|
||||
|
||||
@@ -5,4 +5,5 @@ python-multipart==0.0.9
|
||||
pytest==8.3.3
|
||||
httpx==0.27.2
|
||||
cryptography==43.0.1
|
||||
Pillow==11.3.0
|
||||
# stdlib only for sqlite3, logging, json, contextvars, threading
|
||||
|
||||
@@ -126,3 +126,82 @@ def test_test_inference_unconfigured(env):
|
||||
r = client.post("/api/v1/admin/test-inference")
|
||||
assert r.status_code == 200
|
||||
assert r.json()["ok"] is False and r.json()["mode"] == "fail"
|
||||
|
||||
|
||||
def _rgba_png(w: int, h: int) -> bytes:
|
||||
"""RGBA PNG: opaque body + transparent top-left block (capture-like alpha)."""
|
||||
from PIL import Image
|
||||
import io
|
||||
im = Image.new("RGBA", (w, h), (10, 20, 30, 255))
|
||||
for y in range(min(80, h)):
|
||||
for x in range(min(80, w)):
|
||||
im.putpixel((x, y), (0, 0, 0, 0))
|
||||
buf = io.BytesIO()
|
||||
im.save(buf, format="PNG")
|
||||
return buf.getvalue()
|
||||
|
||||
|
||||
def test_fit_long_edge_png_downscales_over_1024():
|
||||
"""METADATA-35: Call 1 images long edge <= 1024; never upscale; keep alpha."""
|
||||
from backend.inference import fit_long_edge_png
|
||||
from PIL import Image
|
||||
import io
|
||||
|
||||
src = _rgba_png(800, 1067) # inventory 3:4, long edge 1067
|
||||
out = fit_long_edge_png(src, max_long_edge=1024)
|
||||
im = Image.open(io.BytesIO(out)).convert("RGBA")
|
||||
assert max(im.size) == 1024
|
||||
assert min(im.size) < 1024
|
||||
w, h = im.size
|
||||
assert abs((w / h) - (800 / 1067)) < 0.02
|
||||
# Transparent block still transparent near origin after downscale
|
||||
assert im.getpixel((0, 0))[3] == 0
|
||||
assert im.getpixel((5, 5))[3] == 0
|
||||
|
||||
|
||||
def test_fit_long_edge_png_leaves_small_unchanged_dims():
|
||||
from backend.inference import fit_long_edge_png
|
||||
from PIL import Image
|
||||
import io
|
||||
|
||||
src = _rgba_png(400, 300)
|
||||
out = fit_long_edge_png(src, max_long_edge=1024)
|
||||
im = Image.open(io.BytesIO(out))
|
||||
assert im.size == (400, 300)
|
||||
|
||||
|
||||
def test_call1_sends_long_edge_at_most_1024(env, monkeypatch):
|
||||
"""METADATA-35 end-to-end: messages to the model use <=1024 long edge."""
|
||||
client, conn = env
|
||||
conn.execute("INSERT OR REPLACE INTO settings(key,value,value_enc) VALUES ('inference_enabled','true',NULL)")
|
||||
conn.execute(
|
||||
"INSERT INTO slabs(id,sku,length_in,thickness_in,status,server_rev,client_rev,created_at,updated_at) "
|
||||
"VALUES ('s1','SLAB-01',100,2,'calibrated',1,1,'t','t')"
|
||||
)
|
||||
big = _rgba_png(800, 1067)
|
||||
conn.execute(
|
||||
"INSERT INTO slab_photos(id,slab_id,seq,processed_blob,created_at) VALUES ('p1','s1',1,?, 't')",
|
||||
(big,),
|
||||
)
|
||||
conn.commit()
|
||||
from backend.inference import InferenceClient
|
||||
from PIL import Image
|
||||
import io
|
||||
|
||||
captured = {}
|
||||
|
||||
def fake_chat(self, messages):
|
||||
captured["messages"] = messages
|
||||
return json.dumps({"species_id": 1, "species_confidence": 0.9})
|
||||
|
||||
monkeypatch.setattr(InferenceClient, "chat", fake_chat)
|
||||
r = client.post("/api/v1/slabs/s1/inference/call1")
|
||||
assert r.status_code == 200, r.text
|
||||
content = captured["messages"][0]["content"]
|
||||
imgs = [p for p in content if isinstance(p, dict) and p.get("type") == "image_url"]
|
||||
assert len(imgs) == 1
|
||||
url = imgs[0]["image_url"]["url"]
|
||||
assert url.startswith("data:image/png;base64,")
|
||||
raw = base64.b64decode(url.split(",", 1)[1])
|
||||
im = Image.open(io.BytesIO(raw))
|
||||
assert max(im.size) <= 1024
|
||||
|
||||
@@ -63,7 +63,7 @@ The listing is built in two inference calls, both optional. Call 1 (vision) read
|
||||
| METADATA-25 | Species, character, length, and free-text thickness are not product attributes. | M3 | — |
|
||||
| METADATA-30 | Product tags come from the synced Woo set and are all optional; no hardcoded prefixes, and the bare fig tag is kept. | M3 | — |
|
||||
| METADATA-100 | Call 1 free-text observations (character, inclusions, voids, checks) are mapped to existing synced feat-* tags only; WooCommerce is the authority for feat-* definitions. | M3 | — |
|
||||
| METADATA-35 | Call 1 runs only when inference is enabled, automatically after a calibrated draft uploads or on demand, and sends the clean processed photos at 1024 px long edge or less plus user metadata and the synced taxonomy. | M3 | — |
|
||||
| METADATA-35 | Call 1 runs only when inference is enabled, automatically after a calibrated draft uploads or on demand, and sends the clean processed photos at 1024 px long edge or less plus user metadata and the synced taxonomy. | M3 | backend/tests/test_inference.py |
|
||||
| METADATA-40 | Call 1 returns species, wood categories, edge, figure, and feature-tag suggestions constrained to the synced lists, each with confidence. Grade is not suggested by Call 1 (owner judgment). | M3 | — |
|
||||
| METADATA-45 | Call 1 confidence at or above the threshold (default 0.7) pre-fills review fields; below it the field stays empty for the owner to fill by hand or re-run Call 1; suggestions are always editable. | M3 | — |
|
||||
| METADATA-50 | Typed values always win over inferred values. | M3 | — |
|
||||
|
||||
Reference in New Issue
Block a user