Spaces:
Sleeping
Sleeping
File size: 40,513 Bytes
be197b9 7b374c9 be197b9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 | # %%
# ============================================================
# JEWELLERY MULTIMODAL SEARCH BACKEND (FASTAPI)
# ============================================================
# %%
# ============================================================
# IMPORTS
# ============================================================
import os
import json
from typing import List, Dict
import torch
import clip
import numpy as np
import chromadb
from fastapi import FastAPI, HTTPException, File, UploadFile
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import FileResponse
from pydantic import BaseModel
from openai import OpenAI
from dotenv import load_dotenv
from sentence_transformers import CrossEncoder
import base64
import requests
from PIL import Image
import io
# Load environment variables from .env file
load_dotenv()
# %%
# ============================================================
# CONFIG
# ============================================================
# Use absolute paths for deployment
# Priority: /app (Docker/HF Spaces) > script directory (local)
if os.path.exists("/app"):
# Running in Docker (Hugging Face Spaces)
BASE_DIR = "/app"
else:
# Running locally - use absolute path to script directory
BASE_DIR = os.path.abspath(os.path.dirname(__file__))
CHROMA_PATH = os.path.join(BASE_DIR, "chroma_primary")
DATA_DIR = os.path.join(BASE_DIR, "data", "tanishq")
IMAGE_DIR = os.path.join(DATA_DIR, "images")
BLIP_CAPTIONS_PATH = os.path.join(DATA_DIR, "blip_captions.json")
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
# %%
# ============================================================
# LAZY MODEL LOADING (Reduces cold start time)
# ============================================================
# Global model references (loaded on first use)
clip_model = None
cross_encoder = None
def get_clip_model():
"""Lazy load CLIP model on first use"""
global clip_model
if clip_model is None:
print("πΉ Loading CLIP model...")
model, _ = clip.load("ViT-B/16", device=DEVICE)
model.eval()
clip_model = model
print("β
CLIP model loaded")
return clip_model
def get_cross_encoder():
"""Lazy load Cross-Encoder on first use"""
global cross_encoder
if cross_encoder is None:
print("πΉ Loading Cross-Encoder for re-ranking...")
cross_encoder = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
print("β
Cross-Encoder loaded")
return cross_encoder
# %%
print("πΉ Loading BLIP captions...")
with open(BLIP_CAPTIONS_PATH, "r") as f:
BLIP_CAPTIONS = json.load(f)
# %%
# ============================================================
# INITIALIZE GROQ LLM CLIENT
# ============================================================
GROQ_API_KEY = os.environ.get("GROQ_API_KEY")
if GROQ_API_KEY:
print("πΉ Initializing Groq LLM client...")
groq_client = OpenAI(
base_url="https://api.groq.com/openai/v1",
api_key=GROQ_API_KEY
)
else:
groq_client = None
print("β οΈ GROQ_API_KEY not set; LLM features disabled (fallbacks enabled)")
# NVIDIA OCR API configuration
NVIDIA_API_KEY = os.environ.get("NVIDIA_API_KEY")
NVIDIA_OCR_URL = "https://ai.api.nvidia.com/v1/cv/nvidia/nemoretriever-ocr-v1"
# Fallback OCR configuration
OPENAI_API_KEY = os.environ.get("OPENAI_API_KEY")
# %%
# ============================================================
# LOAD CHROMA (PERSISTED DB)
# ============================================================
print("πΉ Connecting to Chroma DB...")
chroma_client = chromadb.PersistentClient(path=CHROMA_PATH)
image_collection = chroma_client.get_collection("jewelry_images")
metadata_collection = chroma_client.get_collection("jewelry_metadata")
print(
"β
Chroma loaded | Images:",
image_collection.count(),
"| Metadata:",
metadata_collection.count()
)
# %%
# ============================================================
# FASTAPI APP
# ============================================================
app = FastAPI(title="Jewellery Multimodal Search")
@app.get("/")
def root():
return {
"status": "Jewellery Search API running",
"docs": "/docs",
"health": "/health"
}
app.add_middleware(
CORSMiddleware,
allow_origins=[
"http://localhost:5173",
"https://tanishq-rag-capstone-1lj2x4y1v-akash-aimls-projects.vercel.app", # Vercel preview
"https://*.vercel.app", # All Vercel deployments
"https://*.ngrok-free.app", # Allow ngrok tunnels
"https://*.ngrok-free.dev", # Allow ngrok tunnels (new domain)
"https://*.ngrok.io", # Allow ngrok tunnels (legacy)
],
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"],
)
# %%
# ============================================================
# MIDDLEWARE FOR HF SPACES OPTIMIZATION
# ============================================================
import asyncio
from starlette.requests import Request
@app.middleware("http")
async def add_optimizations(request: Request, call_next):
"""Add upload size limits and request timeouts"""
# Limit upload size to 5MB
if request.method == "POST":
content_length = request.headers.get("content-length")
if content_length and int(content_length) > 5 * 1024 * 1024: # 5MB
raise HTTPException(status_code=413, detail="File too large (max 5MB)")
# Add request timeout (60s for local dev/slower machines)
try:
response = await asyncio.wait_for(call_next(request), timeout=60.0)
return response
except asyncio.TimeoutError:
raise HTTPException(status_code=504, detail="Request timeout (max 60s)")
# %%
# ============================================================
# REQUEST / RESPONSE SCHEMAS
# ============================================================
class TextSearchRequest(BaseModel):
query: str
filters: Dict[str, str] = None # Explicit UI filters (e.g. {"metal": "gold"})
top_k: int = 5
use_reranking: bool = True # Toggle cross-encoder (3x faster when False)
use_explanations: bool = True # Toggle LLM explanations (500ms+ faster when False)
class SimilarSearchRequest(BaseModel):
image_id: str
top_k: int = 5
# %%
# ============================================================
# CLIP QUERY ENCODING (TEXT ONLY)
# ============================================================
def encode_text_clip(text: str) -> np.ndarray:
"""Encode text using CLIP with memory cleanup"""
model = get_clip_model()
tokens = clip.tokenize([text]).to(DEVICE)
with torch.no_grad():
emb = model.encode_text(tokens)
emb = emb / emb.norm(dim=-1, keepdim=True)
result = emb.cpu().numpy()[0]
# Memory cleanup
del tokens, emb
if DEVICE == "cuda":
torch.cuda.empty_cache()
return result
# %%
# ============================================================
# INTENT & ATTRIBUTE DETECTION WITH LLM (STRUCTURED)
# ============================================================
def detect_intent_and_attributes(query: str) -> Dict:
"""
Extract search attributes and exclusions from query using LLM with fixed schema.
Returns:
{
"intent": "search",
"attributes": {category, metal, primary_stone}, # Items to INCLUDE
"exclusions": {category, metal, primary_stone} # Items to EXCLUDE
}
"""
prompt = f"""Extract jewellery search attributes from this query.
Query: "{query}"
Return ONLY valid JSON with this exact schema:
{{
"intent": "search",
"attributes": {{
"category": "ring|necklace|earring|bracelet|null",
"metal": "gold|silver|platinum|null",
"primary_stone": "diamond|pearl|ruby|emerald|sapphire|null"
}},
"exclusions": {{
"category": "ring|necklace|earring|bracelet|null",
"metal": "gold|silver|platinum|null",
"primary_stone": "diamond|pearl|ruby|emerald|sapphire|null"
}}
}}
Rules:
- "attributes" = what to INCLUDE (positive filters)
- "exclusions" = what to EXCLUDE (negative filters)
- Use null for unspecified fields
- Detect negations: "no", "without", "not", "plain", "-free"
Examples:
Query: "gold ring with diamonds"
{{"intent": "search", "attributes": {{"category": "ring", "metal": "gold", "primary_stone": "diamond"}}, "exclusions": {{"category": null, "metal": null, "primary_stone": null}}}}
Query: "ring with no diamonds"
{{"intent": "search", "attributes": {{"category": "ring", "metal": null, "primary_stone": null}}, "exclusions": {{"category": null, "metal": null, "primary_stone": "diamond"}}}}
Query: "plain silver necklace"
{{"intent": "search", "attributes": {{"category": "necklace", "metal": "silver", "primary_stone": null}}, "exclusions": {{"category": null, "metal": null, "primary_stone": "any"}}}}
Query: "gold necklace without pearls"
{{"intent": "search", "attributes": {{"category": "necklace", "metal": "gold", "primary_stone": null}}, "exclusions": {{"category": null, "metal": null, "primary_stone": "pearl"}}}}
Return ONLY the JSON, no explanation."""
def simple_fallback() -> Dict:
q = query.lower()
attrs = {}
if "necklace" in q:
attrs["category"] = "necklace"
elif "ring" in q:
attrs["category"] = "ring"
if "gold" in q:
attrs["metal"] = "gold"
elif "silver" in q:
attrs["metal"] = "silver"
if "pearl" in q:
attrs["primary_stone"] = "pearl"
elif "diamond" in q:
attrs["primary_stone"] = "diamond"
return {
"intent": "search",
"attributes": attrs,
"exclusions": {}
}
if groq_client is None:
return simple_fallback()
try:
response = groq_client.chat.completions.create(
model="llama-3.3-70b-versatile",
messages=[{"role": "user", "content": prompt}],
temperature=0.1,
max_tokens=300
)
result_text = response.choices[0].message.content.strip()
# Extract JSON from response (handle markdown code blocks)
if "```json" in result_text:
result_text = result_text.split("```json")[1].split("```")[0].strip()
elif "```" in result_text:
result_text = result_text.split("```")[1].split("```")[0].strip()
result = json.loads(result_text)
# Clean null values
result["attributes"] = {k: v for k, v in result.get("attributes", {}).items() if v and v != "null"}
result["exclusions"] = {k: v for k, v in result.get("exclusions", {}).items() if v and v != "null"}
return result
except Exception as e:
print(f"β οΈ LLM extraction failed: {e}, falling back to simple extraction")
return simple_fallback()
# %%
# ============================================================
# VISUAL RETRIEVAL (NO LANGCHAIN)
# ============================================================
def retrieve_visual_candidates(query_text: str, k: int = 100, where_filter: Dict = None):
q_emb = encode_text_clip(query_text)
# Use Chroma's built-in filtering if provided
res = image_collection.query(
query_embeddings=[q_emb],
n_results=k,
where=where_filter
)
if not res["ids"] or not res["ids"][0]:
return []
return [
{
"image_id": img_id,
"visual_score": dist
}
for img_id, dist in zip(res["ids"][0], res["distances"][0])
]
# %%
# ============================================================
# METADATA SCORING REFINEMENTS
# ============================================================
def adaptive_alpha(query_attrs: Dict) -> float:
return 0.1 + 0.1 * len(query_attrs)
def refined_metadata_adjustment(meta: Dict, query_attrs: Dict) -> float:
score = 0.0
for attr, q_val in query_attrs.items():
m_val = meta.get(attr)
conf = meta.get(f"confidence_{attr}", 0.0)
if m_val == q_val:
score += conf
elif conf > 0.6:
score -= 0.3 * conf
return score
def apply_metadata_boost(candidates: List[Dict], query_attrs: Dict, exclusions: Dict = None):
"""
Rank candidates by combining visual similarity with metadata matching.
HARD FILTER out excluded items completely.
Args:
candidates: List of {image_id, visual_score}
query_attrs: Attributes to INCLUDE (boost matching items)
exclusions: Attributes to EXCLUDE (FILTER OUT completely)
"""
if exclusions is None:
exclusions = {}
alpha = adaptive_alpha(query_attrs)
ranked = []
for c in candidates:
meta = metadata_collection.get(
ids=[c["image_id"]],
include=["metadatas"]
)["metadatas"][0]
# HARD FILTER: Skip items that match exclusions
should_exclude = False
for attr, excluded_value in exclusions.items():
meta_value = meta.get(attr)
# Handle "any" exclusion (e.g., "plain" means no stones at all)
if excluded_value == "any":
# Exclude if has ANY stone (not unknown/null)
if meta_value and meta_value not in ["unknown", "null", ""]:
should_exclude = True
print(f"π« Excluding {c['image_id']}: has {attr}={meta_value} (want none)")
break
# Handle specific exclusion
elif meta_value == excluded_value:
should_exclude = True
print(f"π« Excluding {c['image_id']}: has {attr}={meta_value} (excluded)")
break
# Skip this item if it matches any exclusion
if should_exclude:
continue
# Calculate positive boost from matching attributes
adjust = refined_metadata_adjustment(meta, query_attrs)
# Final score: visual + metadata boost (no exclusion penalty needed)
final_score = c["visual_score"] - alpha * adjust
ranked.append({
"image_id": c["image_id"],
"visual_score": c["visual_score"],
"metadata_boost": adjust,
"final_score": final_score
})
return sorted(ranked, key=lambda x: x["final_score"])
def rerank_with_cross_encoder(
query: str,
candidates: List[Dict],
top_k: int = 12
) -> List[Dict]:
"""
Re-rank candidates using cross-encoder for better semantic matching.
Two-stage pipeline:
1. CLIP bi-encoder: Fast retrieval (already done)
2. Cross-encoder: Accurate semantic re-ranking
Args:
query: User query text
candidates: List of {image_id, visual_score, metadata_boost, ...}
top_k: Number of results to return
Returns:
Re-ranked list of top K candidates
"""
if not candidates:
return []
# Prepare query-document pairs for cross-encoder
pairs = []
for c in candidates:
# Get BLIP caption for this image
caption = BLIP_CAPTIONS.get(c["image_id"], "")
# Get metadata
meta = metadata_collection.get(
ids=[c["image_id"]],
include=["metadatas"]
)["metadatas"][0]
# Create rich text representation combining caption + metadata
doc_text = f"{caption}. Category: {meta.get('category', 'unknown')}, Metal: {meta.get('metal', 'unknown')}, Stone: {meta.get('primary_stone', 'unknown')}"
pairs.append([query, doc_text])
# Score all pairs with cross-encoder (batch processing)
print(f"π Cross-encoder scoring {len(pairs)} candidates...")
encoder = get_cross_encoder()
cross_scores = encoder.predict(pairs, batch_size=32)
# Combine scores: visual + metadata + cross-encoder
for i, c in enumerate(candidates):
c["cross_encoder_score"] = float(cross_scores[i])
# Final score combines all signals
# - Visual similarity (CLIP): 30% weight
# - Metadata match: 20% weight
# - Semantic similarity (cross-encoder): 50% weight (highest)
c["final_score_reranked"] = (
-c["visual_score"] * 0.3 + # Negate because lower distance = better
c.get("metadata_boost", 0) * 0.2 +
c["cross_encoder_score"] * 0.5
)
# Sort by final score (higher is better)
ranked = sorted(candidates, key=lambda x: x["final_score_reranked"], reverse=True)
print(f"β
Re-ranked {len(ranked)} candidates, returning top {top_k}")
return ranked[:top_k]
# %%
# ============================================================
# IMAGE UPLOAD & OCR HELPER FUNCTIONS
# ============================================================
def encode_uploaded_image(image_bytes: bytes) -> np.ndarray:
"""Encode uploaded image using CLIP model"""
try:
# Open image from bytes
image = Image.open(io.BytesIO(image_bytes))
# Convert to RGB if necessary
if image.mode != 'RGB':
image = image.convert('RGB')
# Resize if too large (max 512x512 for efficiency)
max_size = 512
if max(image.size) > max_size:
image.thumbnail((max_size, max_size), Image.Resampling.LANCZOS)
# Preprocess for CLIP
from torchvision.transforms import Compose, Resize, CenterCrop, ToTensor, Normalize
preprocess = Compose([
Resize(224, interpolation=Image.BICUBIC),
CenterCrop(224),
ToTensor(),
Normalize((0.48145466, 0.4578275, 0.40821073),
(0.26862954, 0.26130258, 0.27577711))
])
image_tensor = preprocess(image).unsqueeze(0).to(DEVICE)
# Encode with CLIP
model = get_clip_model()
with torch.no_grad():
image_features = model.encode_image(image_tensor)
image_features = image_features / image_features.norm(dim=-1, keepdim=True)
result = image_features.cpu().numpy()[0]
# Memory cleanup
del image_tensor, image_features
if DEVICE == "cuda":
torch.cuda.empty_cache()
return result
except Exception as e:
raise HTTPException(status_code=400, detail=f"Failed to process image: {str(e)}")
def extract_text_from_image(image_bytes: bytes) -> str:
"""Extract text from image using NVIDIA NeMo Retriever OCR API with GPT-4.1-Nano fallback"""
# Try NVIDIA OCR first if key is configured
extracted_text = ""
nvidia_failed = False
if NVIDIA_API_KEY:
try:
# Encode image to base64
image_b64 = base64.b64encode(image_bytes).decode()
# Prepare request
headers = {
"Authorization": f"Bearer {NVIDIA_API_KEY}",
"Accept": "application/json",
"Content-Type": "application/json"
}
payload = {
"input": [
{
"type": "image_url",
"url": f"data:image/png;base64,{image_b64}"
}
]
}
# Call NVIDIA OCR API
print(f"π Calling NVIDIA OCR API...")
response = requests.post(
NVIDIA_OCR_URL,
headers=headers,
json=payload,
timeout=15 # Shorter timeout to fail fast
)
if response.status_code == 200:
result = response.json()
# Format 1: Text detections array
if "data" in result and isinstance(result["data"], list) and len(result["data"]) > 0:
for data_item in result["data"]:
if isinstance(data_item, dict) and "text_detections" in data_item:
for detection in data_item["text_detections"]:
if "text_prediction" in detection and "text" in detection["text_prediction"]:
extracted_text += detection["text_prediction"]["text"] + " "
elif isinstance(data_item, dict) and "content" in data_item:
extracted_text += data_item["content"] + " "
# Format 2: Direct text field
elif "text" in result:
extracted_text = result["text"]
# Format 3: Choices/Results
elif "choices" in result and len(result["choices"]) > 0:
if "text" in result["choices"][0]:
extracted_text = result["choices"][0]["text"]
elif "message" in result["choices"][0]:
extracted_text = result["choices"][0]["message"].get("content", "")
extracted_text = extracted_text.strip()
if extracted_text:
print(f"β
Extracted text (NVIDIA): '{extracted_text}'")
return extracted_text
print(f"β οΈ NVIDIA OCR failed with status {response.status_code}. Trying fallback...")
nvidia_failed = True
except Exception as e:
print(f"β οΈ NVIDIA OCR exception: {e}. Trying fallback...")
nvidia_failed = True
else:
print("βΉοΈ NVIDIA_API_KEY not set. Using fallback directly.")
nvidia_failed = True
# FALLBACK: Custom GPT-4.1-Nano OCR (OpenAI Compatible)
try:
if not OPENAI_API_KEY:
raise HTTPException(status_code=500, detail="OCR unavailable: Primary failed and OPENAI_API_KEY not set for fallback.")
print("π Using Global GPT-4.1-Nano Fallback...")
# Initialize OpenAI client with custom base URL
from openai import OpenAI
client = OpenAI(
base_url="https://apidev.navigatelabsai.com/v1",
api_key=OPENAI_API_KEY
)
image_b64 = base64.b64encode(image_bytes).decode()
response = client.chat.completions.create(
model="gpt-4.1-nano",
messages=[
{
"role": "user",
"content": [
{"type": "text", "text": "Transcribe the handwritten text in this image exactly as it appears. Output ONLY the text, nothing else."},
{
"type": "image_url",
"image_url": {
"url": f"data:image/png;base64,{image_b64}"
}
}
]
}
],
max_tokens=300
)
extracted_text = response.choices[0].message.content.strip()
if not extracted_text:
raise HTTPException(status_code=400, detail="No readable text found in image (Fallback).")
print(f"β
Extracted text (Fallback GPT): '{extracted_text}'")
return extracted_text
except HTTPException:
raise
except Exception as e:
print(f"β Fallback OCR failed: {e}")
raise HTTPException(status_code=500, detail=f"OCR failed permanently: {str(e)}")
# %%
# ============================================================
# LLM-POWERED EXPLANATION (GROQ LLAMA 3.1) - BATCH PROCESSING
# ============================================================
def batch_generate_explanations(results: List[Dict], query_attrs: Dict, user_query: str) -> List[str]:
"""Generate diverse, LLM-powered explanations for all search results in ONE API call"""
if not results:
return []
# Build context for all items (handle up to 20 items in one call)
items_context = []
for idx, r in enumerate(results, 1):
meta = metadata_collection.get(
ids=[r["image_id"]],
include=["metadatas"]
)["metadatas"][0]
matched_attrs = [v for k, v in query_attrs.items() if meta.get(k) == v]
# Compact format to save tokens
item_info = f"{idx}. {meta.get('category', 'item')} | {meta.get('metal', '?')} | {meta.get('primary_stone', '?')} | score:{r['visual_score']:.2f} | matched:{','.join(matched_attrs) if matched_attrs else 'none'}"
items_context.append(item_info)
# Compact prompt to fit more items
prompt = f"""Query: "{user_query}"
Write 1 brief sentence for EACH item:
{chr(10).join(items_context)}
Format:
1. [sentence]
2. [sentence]
etc."""
if groq_client is None:
explanations = []
else:
try:
# Single API call for ALL items
response = groq_client.chat.completions.create(
model="llama-3.1-8b-instant",
messages=[
{"role": "system", "content": "Write brief jewellery recommendations."},
{"role": "user", "content": prompt}
],
temperature=0.7,
max_tokens=min(800, len(results) * 60), # Increased for 12+ items
top_p=0.9
)
# Parse response
full_response = response.choices[0].message.content.strip()
explanations = []
import re
pattern = r'^\s*(\d+)[\.:)\-]\s*(.+?)(?=^\s*\d+[\.:)\-]|\Z)'
matches = re.findall(pattern, full_response, re.MULTILINE | re.DOTALL)
if matches and len(matches) >= len(results):
for num, text in matches[:len(results)]:
clean_text = ' '.join(text.strip().split())
if clean_text and len(clean_text) > 10:
explanations.append(clean_text)
if len(explanations) >= len(results):
return explanations[:len(results)]
# If incomplete, pad with fallback
print(f"β οΈ LLM returned {len(explanations)}/{len(results)} explanations, padding with fallback")
except Exception as e:
print(f"β οΈ LLM explanation failed: {e}, using fallback")
explanations = []
# Fallback for missing explanations
while len(explanations) < len(results):
idx = len(explanations)
r = results[idx]
meta = metadata_collection.get(
ids=[r["image_id"]],
include=["metadatas"]
)["metadatas"][0]
matched_attrs = [v for k, v in query_attrs.items() if meta.get(k) == v]
category = meta.get('category', 'item')
metal = meta.get('metal', 'unknown')
stone = meta.get('primary_stone', 'unknown')
if matched_attrs and r['visual_score'] < 1.3:
explanations.append(
f"Excellent {category} featuring {' and '.join(matched_attrs)}. High visual similarity (score: {r['visual_score']:.2f})."
)
elif matched_attrs:
explanations.append(
f"Beautiful {metal} {category} with {stone}. Features {' and '.join(matched_attrs)}."
)
elif r['visual_score'] < 1.3:
explanations.append(
f"Highly similar {category} with excellent visual match. {metal.capitalize()} with {stone}."
)
else:
explanations.append(
f"Recommended {metal} {category} with {stone}. Good visual similarity."
)
return explanations
# ============================================================
# API ENDPOINTS
# ============================================================
@app.get("/health")
def health_check():
"""Health check endpoint for HF Spaces monitoring"""
return {
"status": "healthy",
"models_loaded": {
"clip": clip_model is not None,
"cross_encoder": cross_encoder is not None,
"blip_captions": len(BLIP_CAPTIONS) > 0
},
"database": {
"images": image_collection.count(),
"metadata": metadata_collection.count()
}
}
@app.post("/search/text")
def search_text(req: TextSearchRequest):
# Detect intent from text
if req.query.strip():
intent = detect_intent_and_attributes(req.query)
attrs = intent["attributes"]
else:
intent = {"intent": "filter", "attributes": {}, "exclusions": {}}
attrs = {}
# === DUAL-STAGE FILTERING STRATEGY ===
# 1. Identify valid IDs from Metadata Collection (Source of Truth)
# 2. Use those IDs to filter Vector Search results
# Construct WHERE clause for Metadata Collection
where_clauses = []
if req.filters:
for key, value in req.filters.items():
where_clauses.append({key: value.lower()}) # Explicit filters
# Also apply attributes detected from text as generic filters if user didn't specify explicit ones
# (Optional: this makes "emerald ring" implies primary_stone=emerald)
# But usually we let visual search handle text unless it's strict.
final_where = None
if len(where_clauses) > 1:
final_where = {"$and": where_clauses}
elif len(where_clauses) == 1:
final_where = where_clauses[0]
valid_ids = None
if final_where:
# Fetch ALL valid IDs matching the filter
print(f"π Filtering metadata with: {final_where}")
meta_res = metadata_collection.get(where=final_where, include=["metadatas"])
if meta_res["ids"]:
valid_ids = set(meta_res["ids"])
print(f"β
Found {len(valid_ids)} valid items matching filters.")
else:
print("β οΈ No items match the filters.")
return {"query": req.query, "intent": attrs, "results": []}
# === EXECUTE SEARCH ===
# Case A: Filter Only (No Text Query)
if not req.query.strip() and valid_ids:
# Just return the matching items (Top K)
candidates = [{"image_id": vid, "visual_score": 0.0} for vid in list(valid_ids)[:req.top_k]]
ranked = candidates # No ranking needed without text
explanations = ["Filtered result"] * len(ranked)
# Case B: Text Query (with or without Filter)
else:
search_query = req.query if req.query.strip() else "jewellery"
# We perform a BROADER vector search, then filter in Python
# Retrieve K*5 or at least 100 to ensure we find intersections
fetch_k = 200 if valid_ids else 40
# Note: We do NOT pass 'where' to retrieve_visual_candidates because
# image_collection lacks metadata. We filter manually.
candidates = retrieve_visual_candidates(search_query, k=fetch_k)
filtered_candidates = []
for c in candidates:
if valid_ids is not None:
if c["image_id"] in valid_ids:
filtered_candidates.append(c)
else:
filtered_candidates.append(c)
# Apply strict limit now
filtered = filtered_candidates # apply_metadata_boost(filtered_candidates, attrs, {})
# Cross-encoder re-ranking
if req.use_reranking and filtered and req.query.strip():
ranked = rerank_with_cross_encoder(req.query, filtered, req.top_k)
else:
ranked = filtered[:req.top_k]
# Explanations
if req.use_explanations and req.query.strip():
explanations = batch_generate_explanations(ranked, attrs, search_query)
else:
explanations = ["Match found"] * len(ranked)
# === FORMAT RESULTS ===
results = []
# Fetch metadata for final results
if ranked:
ranked_ids = [r["image_id"] for r in ranked]
metas = metadata_collection.get(ids=ranked_ids, include=["metadatas"])["metadatas"]
meta_map = {rid: m for rid, m in zip(ranked_ids, metas)}
else:
meta_map = {}
for r, explanation in zip(ranked, explanations):
results.append({
"image_id": r["image_id"],
"explanation": explanation,
"metadata": meta_map.get(r["image_id"], {}),
"scores": {
"visual": r["visual_score"],
"final": r.get("visual_score", 0) # simplified
}
})
return {
"query": req.query,
"intent": attrs,
"results": results
}
return {
"query": req.query,
"intent": attrs,
"results": results
}
# %%
@app.post("/search/similar")
def search_similar(req: SimilarSearchRequest):
base = image_collection.get(
ids=[req.image_id],
include=["embeddings"]
)["embeddings"][0]
res = image_collection.query(
query_embeddings=[base],
n_results=req.top_k + 1
)
base_meta = metadata_collection.get(
ids=[req.image_id],
include=["metadatas"]
)["metadatas"][0]
attrs = {
k: base_meta[k]
for k in ["category", "metal", "primary_stone"]
if base_meta.get(k) != "unknown"
}
candidates = [
{
"image_id": img_id,
"visual_score": dist
}
for img_id, dist in zip(res["ids"][0], res["distances"][0])
if img_id != req.image_id
]
ranked = apply_metadata_boost(candidates, attrs, {})[:req.top_k]
# Generate all explanations in one batch LLM call
# For similar search, use the base image ID as the query context
query_context = f"items similar to {req.image_id}"
explanations = batch_generate_explanations(ranked, attrs, query_context)
results = []
for r, explanation in zip(ranked, explanations):
results.append({
"image_id": r["image_id"],
"explanation": explanation,
"scores": {
"visual": r["visual_score"],
"metadata": r["metadata_boost"],
"final": r["final_score"]
}
})
return {
"base_image": req.image_id,
"results": results
}
# %%
# ============================================================
# IMAGE UPLOAD SEARCH ENDPOINT
# ============================================================
@app.post("/search/upload-image")
async def search_by_uploaded_image(
file: UploadFile = File(...),
top_k: int = 12
):
"""
Search for similar jewellery items by uploading an image.
The image is encoded using CLIP and queried against the database.
"""
# Validate file type
if not file.content_type or not file.content_type.startswith('image/'):
raise HTTPException(status_code=400, detail="File must be an image")
try:
# Read image bytes
image_bytes = await file.read()
# Encode image with CLIP
query_embedding = encode_uploaded_image(image_bytes)
# Query ChromaDB
res = image_collection.query(
query_embeddings=[query_embedding.tolist()],
n_results=min(100, top_k * 10),
include=["distances"]
)
# Get metadata for all results
candidates = []
for img_id, dist in zip(res["ids"][0], res["distances"][0]):
candidates.append({
"image_id": img_id,
"visual_score": dist
})
# Get metadata from first result to infer attributes
if candidates:
base_meta = metadata_collection.get(
ids=[candidates[0]["image_id"]],
include=["metadatas"]
)["metadatas"][0]
attrs = {
k: base_meta[k]
for k in ["category", "metal", "primary_stone"]
if base_meta.get(k) != "unknown"
}
else:
attrs = {}
# Apply metadata boost (no exclusions for image upload)
ranked = apply_metadata_boost(candidates, attrs, {})[:top_k]
# Generate explanations in batch
query_context = f"items visually similar to uploaded image"
explanations = batch_generate_explanations(ranked, attrs, query_context)
results = []
for r, explanation in zip(ranked, explanations):
results.append({
"image_id": r["image_id"],
"explanation": explanation,
"scores": {
"visual": r["visual_score"],
"metadata": r["metadata_boost"],
"final": r["final_score"]
}
})
return {
"query_type": "uploaded_image",
"filename": file.filename,
"results": results
}
except HTTPException:
raise
except Exception as e:
raise HTTPException(status_code=500, detail=f"Image search failed: {str(e)}")
# %%
# ============================================================
# OCR QUERY SEARCH ENDPOINT
# ============================================================
@app.post("/search/ocr-query")
async def search_by_ocr_query(
file: UploadFile = File(...),
top_k: int = 12
):
"""
Extract text from uploaded image using NVIDIA NeMo OCR,
then perform text-based search with the extracted query.
"""
# Validate file type
if not file.content_type or not file.content_type.startswith('image/'):
raise HTTPException(status_code=400, detail="File must be an image")
try:
# Read image bytes
image_bytes = await file.read()
# Extract text using NVIDIA OCR
extracted_text = extract_text_from_image(image_bytes)
print(f"π Extracted text from image: '{extracted_text}'")
# Use the extracted text for normal text search
intent = detect_intent_and_attributes(extracted_text)
attrs = intent["attributes"]
exclusions = intent.get("exclusions", {})
# Stage 1: CLIP retrieval (reduced to k=40 for HF Spaces)
candidates = retrieve_visual_candidates(extracted_text, k=40)
# Stage 2: Metadata boost + exclusion filtering
filtered = apply_metadata_boost(candidates, attrs, exclusions)
# Stage 3: Cross-encoder re-ranking
ranked = rerank_with_cross_encoder(extracted_text, filtered, top_k)
# Generate explanations in batch
explanations = batch_generate_explanations(ranked, attrs, extracted_text)
results = []
for r, explanation in zip(ranked, explanations):
results.append({
"image_id": r["image_id"],
"explanation": explanation,
"scores": {
"visual": r["visual_score"],
"metadata": r["metadata_boost"],
"final": r["final_score"]
}
})
return {
"query_type": "ocr_extracted",
"extracted_text": extracted_text,
"intent": attrs,
"results": results
}
except HTTPException:
raise
except Exception as e:
raise HTTPException(status_code=500, detail=f"OCR search failed: {str(e)}")
# %%
@app.get("/image/{image_id}")
def get_image(image_id: str):
path = os.path.join(IMAGE_DIR, image_id)
if not os.path.exists(path):
raise HTTPException(status_code=404, detail="Image not found")
return FileResponse(path)
# %%
# ============================================================
# RUN SERVER
# ============================================================
if __name__ == "__main__":
import uvicorn
print("π Starting Jewellery Search API server...")
print(f"π Data directory: {DATA_DIR}")
print(f"π Image directory: {IMAGE_DIR}")
print(f"π ChromaDB path: {CHROMA_PATH}")
print(f"π Server will run on: http://localhost:8000")
print(f"π API docs available at: http://localhost:8000/docs")
uvicorn.run(app, host="0.0.0.0", port=8000, log_level="info")
|