MicroHS / app.py
github-actions[bot]
Sync from GitHub 38cd8d69dc858672e22cd1448f7768fef87468b1
79f9b3a
"""
HS Code Classifier Web App
FastAPI backend with:
- Real-time HS code prediction from text input
- Document upload with OCR (Tesseract) support
- Structured field extraction from trade documents
- HS (6-digit) and HTS (7-10 digit) code support
- Top-5 suggestions with confidence scores
- Latent space visualization with UMAP
- Multilingual support (EN, TH, VI, ZH)
"""
import json
import os
import re
import shutil
import tempfile
import threading
import time
import pickle
import uuid
from pathlib import Path
import numpy as np
import pandas as pd
from fastapi import FastAPI, Request, UploadFile, File, Form
from fastapi.responses import HTMLResponse, JSONResponse
from fastapi.staticfiles import StaticFiles
from fastapi.templating import Jinja2Templates
from sentence_transformers import SentenceTransformer
from sklearn.neighbors import KNeighborsClassifier
from sklearn.preprocessing import LabelEncoder
from field_extractor import extract_fields, get_all_countries, get_all_currencies
from hs_dataset import get_dataset, get_hts_extensions, get_available_hts_countries
# Paths
PROJECT_DIR = Path(__file__).parent
MODEL_DIR = PROJECT_DIR / "models"
DATA_DIR = PROJECT_DIR / "data"
UPLOAD_DIR = PROJECT_DIR / "uploads"
UPLOAD_DIR.mkdir(exist_ok=True)
# Upload config
MAX_FILE_SIZE = 10 * 1024 * 1024 # 10MB
ALLOWED_EXTENSIONS = {".png", ".jpg", ".jpeg", ".tiff", ".tif", ".bmp", ".pdf"}
# Initialize FastAPI
from starlette.middleware.gzip import GZipMiddleware
app = FastAPI(title="HS Code Classifier", version="2.0.0")
app.add_middleware(GZipMiddleware, minimum_size=1000)
app.mount("/static", StaticFiles(directory=str(PROJECT_DIR / "static")), name="static")
templates = Jinja2Templates(directory=str(PROJECT_DIR / "templates"))
# Global model state
model = None
classifier = None
label_encoder = None
hs_reference = None
training_data = None
embeddings = None
umap_data = None
umap_ready = False
hs_dataset = None
classifier_training_indices = None
def _download_hf_artifacts():
"""Download large artifacts from HF Hub if not present locally."""
from huggingface_hub import hf_hub_download
repo_id = os.getenv("HF_ARTIFACT_REPO", "Mead0w1ark/multilingual-e5-small-hs-codes")
file_map = {
MODEL_DIR / "embeddings.npy": "embeddings.npy",
MODEL_DIR / "knn_classifier.pkl": "knn_classifier.pkl",
MODEL_DIR / "label_encoder.pkl": "label_encoder.pkl",
MODEL_DIR / "metadata.json": "metadata.json",
MODEL_DIR / "umap_data.json": "umap_data.json",
DATA_DIR / "training_data.csv": "training_data.csv",
}
for local_path, repo_filename in file_map.items():
if not local_path.exists():
print(f"Downloading {repo_filename} from {repo_id}...")
try:
downloaded = hf_hub_download(
repo_id=repo_id, filename=repo_filename,
)
local_path.parent.mkdir(parents=True, exist_ok=True)
shutil.copy2(downloaded, local_path)
print(f" -> {local_path}")
except Exception as e:
print(f" Warning: could not download {repo_filename}: {e}")
def load_models():
"""Load all model artifacts on startup."""
global model, classifier, label_encoder, hs_reference, training_data, embeddings, umap_data, hs_dataset, classifier_training_indices
print("Loading models...")
start = time.time()
# Download large artifacts from HF Hub if missing locally.
_download_hf_artifacts()
# Load sentence transformer:
# prefer local bundled model, fall back to Hub model when large files are not in repo.
local_model_dir = MODEL_DIR / "sentence_model"
has_local_weights = (
(local_model_dir / "model.safetensors").exists()
or (local_model_dir / "pytorch_model.bin").exists()
)
has_local_tokenizer = (local_model_dir / "tokenizer.json").exists()
if local_model_dir.exists() and has_local_weights and has_local_tokenizer:
model = SentenceTransformer(str(local_model_dir))
print("Loaded local sentence model from models/sentence_model")
else:
fallback_model = os.getenv(
"SENTENCE_MODEL_NAME",
"intfloat/multilingual-e5-small",
)
model = SentenceTransformer(fallback_model)
print(f"Loaded sentence model from Hugging Face Hub: {fallback_model}")
# Load HS code reference
with open(DATA_DIR / "hs_codes_reference.json") as f:
hs_reference = json.load(f)
# Load training data
training_data_path = DATA_DIR / "training_data_indexed.csv"
if not training_data_path.exists():
training_data_path = DATA_DIR / "training_data.csv"
training_data = pd.read_csv(training_data_path)
training_data["hs_code"] = training_data["hs_code"].astype(str).str.zfill(6)
classifier_path = MODEL_DIR / "knn_classifier.pkl"
label_encoder_path = MODEL_DIR / "label_encoder.pkl"
embeddings_path = MODEL_DIR / "embeddings.npy"
embeddings_part_paths = sorted(MODEL_DIR.glob("embeddings_part*.npy"))
core_codes = {str(k).zfill(6) for k in hs_reference.keys()}
artifacts_exist = (
classifier_path.exists()
and label_encoder_path.exists()
and (embeddings_path.exists() or len(embeddings_part_paths) > 0)
)
def load_cached_embeddings():
if embeddings_path.exists():
return np.load(embeddings_path)
part_paths = sorted(MODEL_DIR.glob("embeddings_part*.npy"))
if part_paths:
parts = [np.load(p) for p in part_paths]
return np.concatenate(parts, axis=0)
return None
def compute_full_embeddings():
texts = training_data["text"].fillna("").astype(str).tolist()
if not texts:
raise RuntimeError("No training rows available to rebuild classifier.")
return model.encode(
[f"passage: {text}" for text in texts],
normalize_embeddings=True,
convert_to_numpy=True,
)
def rebuild_classifier_on_curated_codes():
global classifier, label_encoder, classifier_training_indices
classifier_df = training_data[training_data["hs_code"].isin(core_codes)].copy()
if classifier_df.empty:
classifier_df = training_data
clf_indices = classifier_df.index.to_numpy()
clf_embeddings = embeddings[clf_indices]
hs_labels = classifier_df["hs_code"].tolist()
label_encoder = LabelEncoder()
y = label_encoder.fit_transform(hs_labels)
classifier = KNeighborsClassifier(
n_neighbors=min(5, len(classifier_df)),
metric="cosine",
weights="distance",
)
classifier.fit(clf_embeddings, y)
classifier_training_indices = clf_indices
print(
f"Rebuilt classifier on {len(classifier_df)} rows "
f"across {len(set(hs_labels))} curated HS codes"
)
try:
np.save(embeddings_path, embeddings)
with open(classifier_path, "wb") as f:
pickle.dump(classifier, f)
with open(label_encoder_path, "wb") as f:
pickle.dump(label_encoder, f)
print("Saved rebuilt classifier artifacts to models/")
except Exception as e:
print(f"Warning: could not cache rebuilt artifacts: {e}")
if artifacts_exist:
with open(classifier_path, "rb") as f:
classifier = pickle.load(f)
with open(label_encoder_path, "rb") as f:
label_encoder = pickle.load(f)
embeddings = load_cached_embeddings()
print("Loaded classifier artifacts from models/")
if embeddings is None or len(embeddings) != len(training_data):
print(
f"Embeddings size mismatch (embeddings={len(embeddings) if embeddings is not None else 0}, "
f"data={len(training_data)}). "
"Recomputing embeddings..."
)
embeddings = compute_full_embeddings()
artifact_codes = {str(c).zfill(6) for c in getattr(label_encoder, "classes_", [])}
invalid_artifacts = (
not artifact_codes
or not artifact_codes.issubset(core_codes)
or len(artifact_codes) > len(core_codes)
)
if invalid_artifacts:
print("Classifier artifacts not aligned with curated HS set; rebuilding classifier...")
rebuild_classifier_on_curated_codes()
else:
# Map KNN fit row indices back to full training_data row indices for latent neighbors.
classifier_df = training_data[training_data["hs_code"].isin(artifact_codes)].copy()
classifier_training_indices = classifier_df.index.to_numpy()
n_fit = int(getattr(classifier, "n_samples_fit_", 0))
if n_fit <= 0:
fit_x = getattr(classifier, "_fit_X", None)
n_fit = int(fit_x.shape[0]) if fit_x is not None else 0
if n_fit > 0 and len(classifier_training_indices) == n_fit:
print(f"Mapped classifier indices to {len(classifier_training_indices)} training rows")
else:
print(
"Classifier index mapping mismatch "
f"(mapped={len(classifier_training_indices)}, fit={n_fit}); rebuilding classifier..."
)
rebuild_classifier_on_curated_codes()
else:
print("Classifier artifacts missing; rebuilding from training data...")
embeddings = compute_full_embeddings()
rebuild_classifier_on_curated_codes()
# Load HS dataset (official harmonized-system data)
hs_dataset = get_dataset()
# UMAP data is loaded/computed in a background thread so the server
# can start immediately and pass the HF Space health check.
umap_data = []
elapsed = time.time() - start
print(f"All models loaded in {elapsed:.1f}s")
def _compute_umap_background():
"""Load UMAP data from cache or compute in background.
Sets the global ``umap_data`` list and ``umap_ready`` flag when done.
"""
global umap_data, umap_ready
cache_path = MODEL_DIR / "umap_data.json"
if cache_path.exists():
try:
with open(cache_path, encoding="utf-8") as f:
cached = json.load(f)
has_category_fields = (
isinstance(cached, list)
and len(cached) > 0
and "chapter_name" in cached[0]
)
if isinstance(cached, list) and len(cached) == len(training_data) and has_category_fields:
umap_data = cached
umap_ready = True
print(f"Loaded cached UMAP data: {len(umap_data)} points")
return
else:
print(
f"Cached UMAP size mismatch (cache={len(cached)}, data={len(training_data)}). "
"Recomputing UMAP projection..."
)
except Exception as e:
print(f"Warning: could not read UMAP cache: {e}")
print("Computing UMAP projection (background)...")
try:
import umap
reducer = umap.UMAP(
n_neighbors=30,
min_dist=0.0,
n_components=2,
metric='cosine',
random_state=42,
)
umap_coords = reducer.fit_transform(embeddings)
points = []
for i, row in training_data.iterrows():
hs_code = str(row["hs_code"]).zfill(6)
chapter = row["hs_chapter"]
chapter_name = str(row.get("hs_chapter_name", "")).strip()
if not chapter_name or re.match(r"^HS\s\d{2}$", chapter_name):
chapter_name = str(chapter).split(";")[0].strip()
desc = hs_reference.get(hs_code, {}).get("desc", "Unknown")
points.append({
"x": float(umap_coords[i, 0]),
"y": float(umap_coords[i, 1]),
"text": row["text"][:80],
"hs_code": hs_code,
"chapter": chapter,
"chapter_name": chapter_name,
"hs_desc": desc,
"language": row["language"],
})
with open(cache_path, "w", encoding="utf-8") as f:
json.dump(points, f, ensure_ascii=False)
umap_data = points
umap_ready = True
print(f"UMAP projection computed for {len(umap_data)} points")
except Exception as e:
print(f"UMAP computation failed: {e}")
umap_ready = True # mark ready so endpoints stop saying "computing"
@app.on_event("startup")
async def startup():
load_models()
threading.Thread(target=_compute_umap_background, daemon=True).start()
@app.get("/", response_class=HTMLResponse)
async def index(request: Request):
"""Main page."""
metadata = {}
try:
with open(MODEL_DIR / "metadata.json") as f:
metadata = json.load(f)
except:
pass
countries = get_all_countries()
currencies = get_all_currencies()
hts_countries = get_available_hts_countries()
return templates.TemplateResponse("index.html", {
"request": request,
"metadata": metadata,
"countries": countries,
"currencies": currencies,
"hts_countries": hts_countries,
})
@app.post("/predict")
async def predict(request: Request):
"""Predict HS code for a product description with optional structured context."""
body = await request.json()
query_text = body.get("text", "").strip()
made_in = body.get("made_in", "")
ship_to = body.get("ship_to", "")
item_price = body.get("item_price", None)
currency = body.get("currency", "")
if not query_text:
return JSONResponse({"error": "No text provided"}, status_code=400)
start = time.time()
# Build enriched query using structured fields
enriched_query = query_text
context_parts = []
if made_in:
context_parts.append(f"origin: {made_in}")
if ship_to:
context_parts.append(f"destination: {ship_to}")
if item_price and currency:
context_parts.append(f"value: {currency} {item_price}")
if context_parts:
enriched_query = f"{query_text} ({', '.join(context_parts)})"
# Encode query with e5 prefix
query_emb = model.encode(
[f"query: {enriched_query}"],
normalize_embeddings=True,
convert_to_numpy=True
)
# Get predictions with probabilities
probs = classifier.predict_proba(query_emb)[0]
top_k = 5
top_indices = np.argsort(probs)[-top_k:][::-1]
predictions = []
for idx in top_indices:
hs_code = label_encoder.classes_[idx]
hs_code_padded = str(hs_code).zfill(6)
confidence = float(probs[idx])
if confidence < 0.01:
continue
info = hs_reference.get(hs_code_padded, {})
chapter_code = hs_code_padded[:2]
heading_code = hs_code_padded[:4]
# Get official description from HS dataset if available
official = hs_dataset.lookup(hs_code_padded) if hs_dataset else None
official_desc = official['description'] if official else None
# Validate against official dataset
validation = hs_dataset.validate_hs_code(hs_code_padded) if hs_dataset else None
predictions.append({
"hs_code": hs_code_padded,
"confidence": confidence,
"description": info.get("desc", official_desc or "No description available"),
"official_description": official_desc,
"chapter": info.get("chapter", "Unknown"),
"chapter_code": chapter_code,
"heading_code": heading_code,
"validated": validation['valid'] if validation else None,
})
# Find nearest training examples
sims = embeddings @ query_emb.T
top_sim_idx = np.argsort(sims.flatten())[-3:][::-1]
similar_examples = []
for idx in top_sim_idx:
if idx < len(training_data):
similar_examples.append({
"text": training_data.iloc[idx]["text"],
"hs_code": str(training_data.iloc[idx]["hs_code"]).zfill(6),
"similarity": float(sims[idx][0]),
})
elapsed = time.time() - start
return JSONResponse({
"query": query_text,
"enriched_query": enriched_query,
"predictions": predictions,
"similar_examples": similar_examples,
"inference_time_ms": round(elapsed * 1000, 1),
})
@app.post("/upload-document")
async def upload_document(file: UploadFile = File(...)):
"""Upload a document (image/PDF) and extract text via OCR + structured fields."""
# Validate file
if not file.filename:
return JSONResponse({"error": "No file provided"}, status_code=400)
ext = Path(file.filename).suffix.lower()
if ext not in ALLOWED_EXTENSIONS:
return JSONResponse(
{"error": f"Unsupported file type: {ext}. Allowed: {', '.join(ALLOWED_EXTENSIONS)}"},
status_code=400
)
# Read file content
content = await file.read()
if len(content) > MAX_FILE_SIZE:
return JSONResponse(
{"error": f"File too large. Maximum: {MAX_FILE_SIZE // (1024*1024)}MB"},
status_code=400
)
# Save to temp file
file_id = str(uuid.uuid4())[:8]
temp_path = UPLOAD_DIR / f"{file_id}{ext}"
with open(temp_path, "wb") as f:
f.write(content)
try:
import pytesseract
from PIL import Image
ocr_text = ""
if ext == ".pdf":
# Convert PDF to images, then OCR
try:
from pdf2image import convert_from_path
images = convert_from_path(str(temp_path), dpi=300)
texts = []
for img in images:
texts.append(pytesseract.image_to_string(img))
ocr_text = "\n\n".join(texts)
except ImportError:
return JSONResponse(
{"error": "PDF support requires pdf2image and poppler. Install with: pip install pdf2image"},
status_code=500
)
except Exception as e:
return JSONResponse(
{"error": f"PDF processing error: {str(e)}"},
status_code=500
)
else:
# Image OCR
img = Image.open(temp_path)
ocr_text = pytesseract.image_to_string(img)
if not ocr_text.strip():
return JSONResponse({
"error": "OCR could not extract any text from this document. Please try a clearer image.",
"raw_text": "",
"fields": {},
})
# Extract structured fields
fields = extract_fields(ocr_text)
return JSONResponse({
"success": True,
"file_id": file_id,
"filename": file.filename,
"raw_text": ocr_text.strip(),
"fields": fields,
})
except Exception as e:
return JSONResponse(
{"error": f"OCR processing failed: {str(e)}"},
status_code=500
)
finally:
# Clean up temp file
if temp_path.exists():
temp_path.unlink()
@app.post("/extract-fields")
async def extract_fields_endpoint(request: Request):
"""Extract structured fields from arbitrary text (no OCR needed)."""
body = await request.json()
text = body.get("text", "").strip()
if not text:
return JSONResponse({"error": "No text provided"}, status_code=400)
fields = extract_fields(text)
return JSONResponse({"fields": fields})
@app.get("/hts-extensions/{hs_code}")
async def get_hts(hs_code: str, country: str = "US"):
"""Get HTS (country-specific) extensions for a 6-digit HS code."""
result = get_hts_extensions(hs_code, country)
return JSONResponse(result)
@app.get("/hs-lookup/{hs_code}")
async def hs_lookup(hs_code: str):
"""Look up an HS code in the official dataset."""
if not hs_dataset:
return JSONResponse({"error": "HS dataset not loaded"}, status_code=500)
result = hs_dataset.lookup(hs_code)
if not result:
# Try search instead
search_results = hs_dataset.search(hs_code, max_results=5)
return JSONResponse({
"found": False,
"message": f"Code {hs_code} not found. Did you mean one of these?",
"suggestions": search_results,
})
return JSONResponse({"found": True, **result})
@app.get("/hs-search")
async def hs_search(q: str = "", limit: int = 20):
"""Search HS codes by description."""
if not q:
return JSONResponse({"error": "No query provided"}, status_code=400)
results = hs_dataset.search(q, max_results=limit)
return JSONResponse({"results": results, "query": q})
@app.get("/hs-validate/{hs_code}")
async def hs_validate(hs_code: str):
"""Validate whether an HS code exists."""
result = hs_dataset.validate_hs_code(hs_code)
return JSONResponse(result)
@app.get("/hts-countries")
async def hts_countries():
"""Get list of countries with HTS extensions available."""
return JSONResponse({"countries": get_available_hts_countries()})
@app.get("/visualization-data")
async def get_visualization_data(request: Request):
"""Return UMAP projection data for visualization.
Supports ``?max_points=N`` to subsample for faster initial load.
The subsample is stratified by chapter so every category is represented.
"""
max_points = int(request.query_params.get("max_points", "0"))
points = umap_data
if not points:
cache_path = MODEL_DIR / "umap_data.json"
if cache_path.exists():
with open(cache_path, encoding="utf-8") as f:
points = json.load(f)
if not points:
if not umap_ready:
return JSONResponse({"points": [], "computing": True})
return JSONResponse({"points": [], "error": "No UMAP data available"})
total = len(points)
if 0 < max_points < total:
# Stratified subsample: keep proportional representation per chapter
import random as _rng
_rng.seed(42)
by_chapter: dict[str, list] = {}
for p in points:
by_chapter.setdefault(p.get("chapter_name", "Other"), []).append(p)
sampled: list = []
for ch, ch_pts in by_chapter.items():
n = max(1, round(len(ch_pts) / total * max_points))
sampled.extend(_rng.sample(ch_pts, min(n, len(ch_pts))))
_rng.shuffle(sampled)
return JSONResponse({"points": sampled, "total": total, "sampled": True})
return JSONResponse({"points": points, "total": total})
@app.get("/visualization-density")
async def get_visualization_density():
"""All UMAP points in compact columnar format for density/labels."""
points = umap_data or []
if not points:
cache_path = MODEL_DIR / "umap_data.json"
if cache_path.exists():
with open(cache_path, encoding="utf-8") as f:
points = json.load(f)
if not points:
if not umap_ready:
return JSONResponse({"chapters": {}, "computing": True})
return JSONResponse({"error": "No data"})
by_chapter: dict[str, dict[str, list]] = {}
for p in points:
ch = p.get("chapter_name", "Unknown")
if ch not in by_chapter:
by_chapter[ch] = {"x": [], "y": []}
by_chapter[ch]["x"].append(round(p["x"], 3))
by_chapter[ch]["y"].append(round(p["y"], 3))
return JSONResponse({"chapters": by_chapter})
@app.post("/embed-query")
async def embed_query(request: Request):
"""Get UMAP coordinates for a query."""
body = await request.json()
query_text = body.get("text", "").strip()
if not query_text:
return JSONResponse({"error": "No text provided"}, status_code=400)
query_emb = model.encode(
[f"query: {query_text}"],
normalize_embeddings=True,
convert_to_numpy=True
)
n_fit = int(getattr(classifier, "n_samples_fit_", 0))
if n_fit <= 0:
fit_x = getattr(classifier, "_fit_X", None)
n_fit = int(fit_x.shape[0]) if fit_x is not None else 0
if n_fit <= 0:
return JSONResponse({"error": "Classifier has no fitted rows"}, status_code=500)
n_neighbors = min(5, n_fit)
distances, indices = classifier.kneighbors(query_emb, n_neighbors=n_neighbors)
if umap_data and len(umap_data) > 0:
weights = 1.0 / (distances[0] + 1e-6)
weights = weights / weights.sum()
mapped_indices = []
for idx in indices[0]:
mapped_idx = int(idx)
if (
classifier_training_indices is not None
and mapped_idx < len(classifier_training_indices)
):
mapped_idx = int(classifier_training_indices[mapped_idx])
mapped_indices.append(mapped_idx)
x = sum(
umap_data[idx]["x"] * w
for idx, w in zip(mapped_indices, weights)
if 0 <= idx < len(umap_data)
)
y = sum(
umap_data[idx]["y"] * w
for idx, w in zip(mapped_indices, weights)
if 0 <= idx < len(umap_data)
)
neighbors = []
for idx, dist in zip(mapped_indices, distances[0]):
if idx < len(umap_data):
point = umap_data[idx]
# cosine distance in [0, 2] for normalized vectors; lower is closer
similarity = max(0.0, min(1.0, 1.0 - float(dist)))
neighbors.append({
**point,
"distance": float(dist),
"similarity": similarity,
})
return JSONResponse({
"x": float(x),
"y": float(y),
"neighbors": neighbors,
})
if not umap_ready:
return JSONResponse({"error": "UMAP data is still computing", "computing": True})
return JSONResponse({"error": "No UMAP data for projection"})
@app.get("/health")
async def health():
"""Health check."""
return {
"status": "ok",
"model_loaded": model is not None,
"hs_dataset_loaded": hs_dataset._loaded if hs_dataset else False,
"hs_codes_count": len(hs_dataset.subheadings) if hs_dataset else 0,
"umap_ready": umap_ready,
}
if __name__ == "__main__":
import uvicorn
uvicorn.run(app, host="0.0.0.0", port=8000)