Update gloss_builder.py

gloss_builder.py

@@ -1,13 +1,128 @@
-from gemma_loader import Gemma3nEdge
+import os
+import requests
+import pandas as pd
 
+def load_known_tokens(csv_path: str = "content/asl_app_data/asl_video_index_final_with_path_cleaned.csv") -> list[str]:
+    """Load vocabulary from CSV — single words only for prompt constraint."""
+    try:
+        df = pd.read_csv(csv_path)
+        tokens = df['token'].dropna().str.upper().str.strip().unique().tolist()
+        # Single words only — more useful for Gemma constraint
+        single = sorted([t for t in tokens if ' ' not in t and len(t) > 1])
+        return single
+    except Exception as e:
+        print(f"⚠️ Could not load tokens from CSV: {e}")
+        # Fallback: words we know we have
+        return [
+            "HELLO", "PLEASE", "COME", "CLEAN", "YOU", "AND", "ANY",
+            "ASK", "ANSWER", "ANOTHER", "ANYONE", "ANYTHING", "ANYWHERE",
+            "APOLOGY", "APPRECIATE", "APPROACH", "ARMY", "AROUND",
+            "ARRIVE", "ART", "ASSIST", "ATTEND", "ATTENTION", "ATTITUDE",
+            "ANGER", "ANIMAL", "ANNOUNCE", "ANNOY", "ANNOYED", "ANXIOUS",
+            "APART", "APARTMENT", "APPLE", "ARROGANT", "ASSOCIATE",
+            "ASSUME", "ATTRACT", "AUNT"
+        ]
 
-def build_gloss_sequence(transcript: str, model_name: str, max_tokens: int = 100) -> list[str]:
+
+def build_gloss_sequence(
+    transcript: str,
+    model_name: str,
+    max_tokens: int = 60,
+    csv_path: str = "content/asl_app_data/asl_video_index_final_with_path_cleaned.csv"
+) -> list[str]:
+    """
+    Convert English transcript to ASL gloss tokens using Gemma.
+    Uses constrained prompt to maximize hits against known vocabulary.
+    Falls back to simple uppercase split if API fails.
+    """
+    known_tokens = load_known_tokens(csv_path)
+    
+    # Take first 300 tokens for prompt — keep it under context limit
+    vocab_sample = ', '.join(known_tokens[:300])
+
+    prompt = f"""You are an expert ASL (American Sign Language) gloss translator.
+
+STRICT RULES:
+1. DROP all articles: a, an, the
+2. DROP all linking verbs: is, are, was, were, am, be, been, being  
+3. DROP all prepositions: to, of, for, in, on, at, by, with, from, into
+4. USE present tense root form only: CLEAN not CLEANING, DELAY not DELAYED, ARRIVE not ARRIVED
+5. ONLY output words from this vocabulary list: {vocab_sample}
+6. If a word has no match, find the CLOSEST SYNONYM from the vocabulary
+7. Output ONLY uppercase tokens separated by spaces — no punctuation, no explanation
+
+EXAMPLES:
+English: "Hello, how are you doing today?"
+ASL Gloss: HELLO YOU
+
+English: "Please clean the dishes"  
+ASL Gloss: PLEASE CLEAN
+
+English: "I am anxious about the appointment"
+ASL Gloss: ANXIOUS APPOINT
+
+English: "Can you come here and help me?"
+ASL Gloss: COME ASSIST
+
+English: "{transcript}"
+ASL Gloss:"""
+
+    api_url = f"https://api-inference.huggingface.co/models/google/gemma-2b-it"
+    headers = {"Authorization": f"Bearer {os.environ.get('HF_TOKEN')}"}
+
+    payload = {
+        "inputs": prompt,
+        "parameters": {
+            "max_new_tokens": max_tokens,
+            "return_full_text": False,
+            "temperature": 0.1,  # Low temp = more deterministic/constrained
+            "do_sample": False
+        }
+    }
+
+    try:
+        response = requests.post(api_url, headers=headers, json=payload, timeout=20)
+        output = response.json()
+
+        raw = ""
+        if isinstance(output, list) and len(output) > 0:
+            raw = output[0].get("generated_text", "").strip()
+        elif isinstance(output, dict):
+            raw = output.get("generated_text", "").strip()
+
+        if not raw:
+            print(f"⚠️ Empty Gemma response, falling back. Raw: {output}")
+            return _simple_gloss(transcript)
+
+        # Clean output — take only first line, strip non-alpha
+        first_line = raw.split('\n')[0].strip()
+        tokens = [
+            t.upper() for t in first_line.split()
+            if t.isalpha() and t.upper() not in {"ASL", "GLOSS", "ENGLISH"}
+        ]
+
+        if not tokens:
+            return _simple_gloss(transcript)
+
+        print(f"✅ Gemma gloss: {tokens}")
+        return tokens
+
+    except Exception as e:
+        print(f"⚠️ Gemma API failed: {e}. Falling back to simple gloss.")
+        return _simple_gloss(transcript)
+
+
+def _simple_gloss(transcript: str) -> list[str]:
     """
-    Returns a list of uppercase ASL gloss tokens using the selected Gemma model.
+    Fallback: strip stopwords and return uppercase tokens.
+    Used when Gemma API is unavailable.
     """
-    # Initialize Gemma with the chosen model directory
-    gemma = Gemma3nEdge(model_dir=model_name)
-    # Generate the raw gloss string
-    raw_gloss = gemma.generate_gloss(transcript, max_tokens=max_tokens)
-    # Split tokens and uppercase for ASL convention
-    return [token.upper() for token in raw_gloss.split()]
+    STOPWORDS = {
+        "A", "AN", "THE", "IS", "ARE", "WAS", "WERE", "AM", "BE", "BEEN",
+        "BEING", "TO", "OF", "FOR", "IN", "ON", "AT", "BY", "WITH", "FROM",
+        "INTO", "AND", "BUT", "OR", "SO", "IF", "AS", "IT", "ITS", "THIS",
+        "THAT", "THESE", "THOSE", "MY", "YOUR", "HIS", "HER", "OUR", "THEIR"
+    }
+    import re
+    words = re.findall(r"[A-Za-z]+", transcript)
+    return [w.upper() for w in words if w.upper() not in STOPWORDS]