Fine-tune NLP model on GAL knowledge base

- Fine-tuned all-mpnet-base-v2 on 898 training pairs from 54 KB topics
- Model uploaded to HF Hub as Clarkoer/gal-mpnet-finetuned
- Similarity scores improved +0.08 to +0.15 across GAL queries
- Updated gal_fallback.py to load fine-tuned model
- Updated Dockerfile to pre-download fine-tuned model
- All 99/99 tests passing

.gitignore

@@ -2,3 +2,5 @@
 __pycache__/
 *.pyc
 .env
+gal-mpnet-finetuned/
+checkpoints/

Backend/gal_fallback.py

@@ -3,7 +3,7 @@ NLP-based fallback responder for the GAL AI chat.
 
 Three-layer hybrid architecture:
   Layer 1 — Rule Engine: regex matches compiler error messages → structured explanations
-  Layer 2 — Retriever: sentence-transformers (all-mpnet-base-v2) semantic search over 50+ KB topics
+  Layer 2 — Retriever: sentence-transformers (gal-mpnet-finetuned) semantic search over 50+ KB topics
   Layer 3 — Default: help menu when nothing matches
 
 Plus: synonym expansion, greeting detection, follow-up context, multi-topic blending.
@@ -704,6 +704,7 @@ Use `prune;` (break) and `skip;` (continue) inside loops."""),
         "if else condition",
         "conditional statement",
         "spring bud wither",
+        "how to use spring bud wither",
         "if statement else if",
         "check a condition",
         "branching logic",
@@ -2222,7 +2223,7 @@ root() {
 
 
 # ═══════════════════════════════════════════════════════════════════════
-# Sentence-Transformers (all-mpnet-base-v2) — lazy-loaded on first query
+# Sentence-Transformers (gal-mpnet-finetuned) — lazy-loaded on first query
 # ═══════════════════════════════════════════════════════════════════════
 
 _st_model = None
@@ -2299,8 +2300,14 @@ def _ensure_model():
         return
 
     from sentence_transformers import SentenceTransformer
-
-    _st_model = SentenceTransformer("all-mpnet-base-v2")
+    import os
+
+    # Prefer fine-tuned model, fall back to base
+    finetuned = os.path.join(os.path.dirname(__file__), "..", "gal-mpnet-finetuned")
+    if os.path.isdir(finetuned):
+        _st_model = SentenceTransformer(finetuned)
+    else:
+        _st_model = SentenceTransformer("Clarkoer/gal-mpnet-finetuned")
 
     _phrase_topic_idx = []
     _responses = []

Dockerfile

@@ -6,8 +6,8 @@ WORKDIR /app
 COPY requirements.txt .
 RUN pip install --no-cache-dir -r requirements.txt
 
-# Pre-download the sentence-transformers model so first request is fast
-RUN python -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('all-mpnet-base-v2')"
+# Pre-download the fine-tuned sentence-transformers model so first request is fast
+RUN python -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('Clarkoer/gal-mpnet-finetuned')"
 
 # Copy the entire project
 COPY . .

train_model.py

@@ -0,0 +1,104 @@
+"""
+Fine-tune all-mpnet-base-v2 on GAL knowledge base for better retrieval.
+
+Strategy: MultipleNegativesRankingLoss with (query, positive) pairs.
+  - Each training phrase is paired with a "topic label" string
+  - Phrases from the same topic should be close, different topics far apart
+
+This produces a fine-tuned model saved to ./gal-mpnet-finetuned/
+"""
+
+import sys, os, random
+sys.path.insert(0, "Backend")
+
+from sentence_transformers import (
+    SentenceTransformer,
+    InputExample,
+    losses,
+    evaluation,
+)
+from torch.utils.data import DataLoader
+
+# ── 1. Load the knowledge base ────────────────────────────────────
+import gal_fallback as fb
+
+topics = []  # list of (phrases, short_label)
+for i, (phrases, response) in enumerate(fb._KNOWLEDGE_BASE):
+    # Create a short label from the first phrase
+    label = phrases[0][:80]
+    topics.append((phrases, label, response))
+
+print(f"Loaded {len(topics)} topics with {sum(len(t[0]) for t in topics)} total phrases")
+
+# ── 2. Generate training pairs ────────────────────────────────────
+# Strategy: for each phrase, pair it with:
+#   a) Other phrases in the same topic (positive pairs)
+#   b) A condensed description of the topic (anchor-passage pair)
+
+train_examples = []
+
+for phrases, label, response in topics:
+    # Take first 200 chars of response as a condensed passage
+    passage = response[:200].strip()
+
+    for phrase in phrases:
+        # Pair each phrase with the passage (query → answer)
+        train_examples.append(InputExample(texts=[phrase, passage]))
+
+        # Pair with other phrases in same topic (query ↔ query)
+        others = [p for p in phrases if p != phrase]
+        if others:
+            partner = random.choice(others)
+            train_examples.append(InputExample(texts=[phrase, partner]))
+
+# Also add synonym-expanded versions
+for term, gal_equiv in fb._SYNONYMS.items():
+    # "what is int" → "what is seed"
+    train_examples.append(
+        InputExample(texts=[f"what is {term} in GAL", f"what is {gal_equiv} in GAL"])
+    )
+    train_examples.append(
+        InputExample(texts=[f"how to use {term}", f"how to use {gal_equiv}"])
+    )
+
+random.shuffle(train_examples)
+print(f"Generated {len(train_examples)} training pairs")
+
+# ── 3. Setup evaluation ──────────────────────────────────────────
+# Build a small eval set: for each topic, take one phrase as query
+# and the response snippet as the expected match
+eval_sentences1 = []
+eval_sentences2 = []
+eval_scores = []
+
+for phrases, label, response in topics:
+    if len(phrases) >= 2:
+        eval_sentences1.append(phrases[-1])  # last phrase as query
+        eval_sentences2.append(response[:200].strip())  # passage
+        eval_scores.append(1.0)  # should be similar
+
+evaluator = evaluation.EmbeddingSimilarityEvaluator(
+    eval_sentences1, eval_sentences2, eval_scores, name="gal-eval"
+)
+
+# ── 4. Fine-tune ─────────────────────────────────────────────────
+model = SentenceTransformer("all-mpnet-base-v2")
+
+train_dataloader = DataLoader(train_examples, shuffle=True, batch_size=16)
+train_loss = losses.MultipleNegativesRankingLoss(model)
+
+output_path = "./gal-mpnet-finetuned"
+
+print("Starting fine-tuning...")
+model.fit(
+    train_objectives=[(train_dataloader, train_loss)],
+    evaluator=evaluator,
+    epochs=5,
+    evaluation_steps=50,
+    warmup_steps=10,
+    output_path=output_path,
+    show_progress_bar=True,
+)
+
+print(f"\nFine-tuned model saved to: {output_path}")
+print("Done!")