update

src/similarity_model/llm_feature_extractor.py

@@ -1,263 +0,0 @@
-import json
-import re
-import time
-
-from src.recommendation_engine.llm_client import generate_text
-
-
-def build_feature_prompt(text: str):
-
-    return f"""
-You are an expert graduation project analyst.
-
-TASK:
-Extract only technical features explicitly mentioned in the project.
-
-RULES:
-- Do not invent anything.
-- Do not generate new ideas.
-- Extract technologies.
-- Extract algorithms.
-- Extract modules.
-- Extract engineering capabilities.
-- Ignore goals and benefits.
-Return between 2 and 10 features.
-
-If fewer than 5 technical features exist,
-return only the available features.
-
-
-
-Technologies such as:
-React,
-Angular,
-Vue,
-MySQL,
-PostgreSQL,
-MongoDB,
-Firebase,
-Django,
-Flask,
-Node.js
-
-must be included if explicitly mentioned.
-
-IMPORTANT:
-Every feature must be copied or paraphrased from the project text.
-Never infer, assume, or generate new capabilities.
-If a capability is not explicitly mentioned, do not include it.
-
-IMPORTANT:
-
-Do not extract:
-- goals
-- benefits
-- outcomes
-- business objectives
-- deployment statements
-
-Return SHORT canonical features.
-
-Maximum 4 words per feature.
-
-Examples:
-computer vision
-deep learning
-internet of things
-route optimization
-waste classification
-
-Do not include long descriptive phrases.
-
-
-IMPORTANT:
-
-Prefer domain-specific capabilities over generic AI terms.
-
-BAD:
-- machine learning
-- deep learning
-- neural networks
-
-GOOD:
-- waste classification
-- route optimization
-- bin fill level monitoring
-- waste generation forecasting
-
-If a specific capability exists, return it instead of the generic AI technology behind it.
-
-Maximum 4 words per feature.
-
-Only extract concrete technical capabilities.
-
-Return JSON only.
-
-Format:
-
-{{
-  "features": [
-    "feature 1",
-    "feature 2"
-  ]
-}}
-
-PROJECT:
-
-{text}
-"""
-
-
-def extract_features_llm(text: str):
-
-    print("CALLING GEMINI")
-
-    prompt = build_feature_prompt(text)
-
-    # Short delay to prevent hitting rate limits
-    time.sleep(1)
-
-    try:
-        response = generate_text(
-            prompt,
-            task="feature",
-            temperature = 0
-        )
-
-        return parse_features_response(
-            response
-        )
-    except Exception as e:
-        print(f"Gemini feature extraction failed, falling back to local dictionary-based keyword matching. Error: {e}")
-        try:
-            import pandas as pd
-            import numpy as np
-            from pathlib import Path
-            
-            possible_paths = [
-                Path(__file__).resolve().parents[2] / "models" / "metadata.parquet",
-                Path("models/metadata.parquet"),
-                Path("../models/metadata.parquet"),
-            ]
-            
-            features_dict = set()
-            for path in possible_paths:
-                if path.exists():
-                    df = pd.read_parquet(path)
-                    for f_list in df['features']:
-                        if isinstance(f_list, (list, np.ndarray, set, tuple)):
-                            features_dict.update([str(f).strip().lower() for f in f_list])
-                    break
-            
-            text_lower = text.lower()
-            matched = []
-            
-            # Sort by length descending to match longer keywords first (e.g. 'deep learning' before 'learning')
-            for f in sorted(list(features_dict), key=len, reverse=True):
-                if len(f) > 2:
-                    pattern = r'\b' + re.escape(f) + r'\b'
-                    if re.search(pattern, text_lower):
-                        matched.append(f)
-                        if len(matched) >= 10:
-                            break
-            print(f"Local fallback matched features: {matched}")
-            return matched
-        except Exception as local_err:
-            print(f"Local feature extraction fallback failed: {local_err}")
-            return []
-
-
-def parse_features_response(response: str):
-
-    if not response:
-        return []
-
-    try:
-
-                # remove markdown json blocks
-        response = response.strip()
-
-        response = re.sub(
-            r"^```(?:json)?",
-            "",
-            response,
-            flags=re.I
-        )
-
-        response = re.sub(
-            r"```$",
-            "",
-            response
-        )
-
-        response = response.strip()
-
-
-        response = response.strip()
-
-        start = response.find("{")
-        end = response.rfind("}")
-
-        if start != -1 and end != -1:
-            response = response[start:end + 1]
-
-        data = json.loads(response)
-
-        features = data.get(
-            "features",
-            []
-        )
-
-        if not isinstance(features, list):
-            return []
-
-        cleaned = []
-
-        NORMALIZATION_MAP = {
-            "computer vision techniques": "computer vision",
-            "deep learning models": "deep learning",
-            "deep neural networks": "deep learning",
-            "machine learning models": "machine learning",
-            "iot sensor networks": "internet of things",
-            "sensor networks": "internet of things",
-            "predictive trends": "predictive analytics",
-            "operational insights": "analytics"
-        }
-
-        for x in features:
-
-            feat = str(x).strip().lower()
-
-            feat = NORMALIZATION_MAP.get(
-                feat,
-                feat
-            )
-
-            if not feat:
-                continue
-
-            if len(feat) < 2:
-                continue
-
-            cleaned.append(feat)
-
-        cleaned = list(dict.fromkeys(cleaned))
-
-        print("PARSED FEATURES:")
-        print(cleaned)
-
-        return cleaned[:10]
-
-    except Exception as e:
-
-        print("=" * 50)
-        print("JSON PARSE ERROR")
-        print(e)
-        print("=" * 50)
-
-        print("RAW RESPONSE:")
-        print(response)
-
-        print("=" * 50)
-
-        return []