Fix: Backend OOM crashes via Vector Cache and worker reduction

Dockerfile

@@ -26,5 +26,5 @@ COPY --chown=user . .
 # Expose port (HF Spaces routes traffic to 7860 by default)
 EXPOSE 7860
 
-# Start Flask — 4 workers so heavy AI requests don't starve auth/ping endpoints
-CMD ["gunicorn", "--bind", "0.0.0.0:7860", "--workers", "4", "--timeout", "120", "--keep-alive", "5", "--max-requests", "500", "--max-requests-jitter", "50", "app:app"]
+# Start Flask — 2 workers is best for 16GB RAM + AI models; 300s timeout for heavy peaks
+CMD ["gunicorn", "--bind", "0.0.0.0:7860", "--workers", "2", "--timeout", "300", "--keep-alive", "5", "--max-requests", "500", "--max-requests-jitter", "50", "app:app"]

model.py

@@ -32,10 +32,12 @@ nli_model = pipeline(
 # RELEVANCE CHECK
 # ==========================================
 def is_relevant(claim_emb, text, threshold=0.15):
+    """Encodes text and checks similarity against claim. 
+    Returns (bool_is_relevant, embedding_as_list)."""
     emb = embed_model.encode([text], normalize_embeddings=True)
     sim = float(np.dot(claim_emb, emb[0]))
     print(f"[DEBUG] Checking relevance for: '{text[:50]}...' Score: {sim:.4f}")
-    return sim >= threshold
+    return sim >= threshold, emb[0].tolist()
 
 def get_search_query(claim):
     stop_words = set(["is", "am", "are", "was", "were", "be", "been", "being",
@@ -72,10 +74,11 @@ def fetch_rss(claim_emb):
             print(f"[RSS] Parsed {url}, found {len(feed.entries)} entries")
             for entry in feed.entries[:5]:
                 title = entry.title
-                # Removed arbitrary 50-char length filter — valid short titles were being dropped
-                if title and is_relevant(claim_emb, title):
-                    save_evidence(title, "RSS")
-                    count += 1
+                if title:
+                    relevant, emb = is_relevant(claim_emb, title)
+                    if relevant:
+                        save_evidence(title, "RSS", embedding=emb)
+                        count += 1
         except Exception as e:
             print(f"[RSS] Error parsing {url}: {e}")
     print(f"[RSS] Saved {count} items.")
@@ -104,10 +107,11 @@ def fetch_gdelt(claim, claim_emb):
 
         for art in articles:
             title = art.get("title", "")
-            # Removed arbitrary 80-char length filter — valid short titles were being dropped
-            if title and is_relevant(claim_emb, title):
-                save_evidence(title, "GDELT")
-                added += 1
+            if title:
+                relevant, emb = is_relevant(claim_emb, title)
+                if relevant:
+                    save_evidence(title, "GDELT", embedding=emb)
+                    added += 1
     except Exception as e:
         print("[WARNING] GDELT failed:", e)
 
@@ -151,9 +155,11 @@ def fetch_newsapi(claim, claim_emb):
             description = art.get("description", "") or ""
             content = f"{title}. {description}".strip(". ")
 
-            if content and is_relevant(claim_emb, content, threshold=0.05):
-                save_evidence(content, f"NewsAPI: {art.get('source', {}).get('name', 'Unknown')}")
-                added += 1
+            if content:
+                relevant, emb = is_relevant(claim_emb, content, threshold=0.05)
+                if relevant:
+                    save_evidence(content, f"NewsAPI: {art.get('source', {}).get('name', 'Unknown')}", embedding=emb)
+                    added += 1
     except Exception as e:
         print("[WARNING] NewsAPI failed:", e)
 
@@ -185,8 +191,9 @@ def fetch_wikipedia(claim):
                 extract = r2.json().get("extract", "")
                 if len(extract) > 20:
                     claim_emb_wiki = embed_model.encode([claim], normalize_embeddings=True)
-                    if is_relevant(claim_emb_wiki, extract, threshold=0.05):
-                        save_evidence(extract, f"Wikipedia: {title}")
+                    relevant, emb = is_relevant(claim_emb_wiki[0], extract, threshold=0.05)
+                    if relevant:
+                        save_evidence(extract, f"Wikipedia: {title}", embedding=emb)
                         saved += 1
         print(f"[Wikipedia] Saved {saved} items.")
 
@@ -204,10 +211,10 @@ def fetch_knowledge_base(claim, claim_emb, threshold=0.30):
     for entry in KNOWLEDGE_BASE:
         text = entry["text"]
         source = entry["source"]
-        emb = embed_model.encode([text], normalize_embeddings=True)
-        sim = float(np.dot(claim_emb, emb[0]))
+        emb_text = embed_model.encode([text], normalize_embeddings=True)
+        sim = float(np.dot(claim_emb, emb_text[0]))
         if sim >= threshold:
-            save_evidence(text, source)
+            save_evidence(text, source, embedding=emb_text[0].tolist())
             saved += 1
     print(f"[KnowledgeBase] Saved {saved} matching entries (threshold={threshold}).")
     return saved
@@ -243,8 +250,9 @@ def fetch_wikidata(claim, claim_emb, threshold=0.10):
             label = item.get("label", "")
             if description and label:
                 text = f"{label}: {description}"
-                if is_relevant(claim_emb, text, threshold=threshold):
-                    save_evidence(text, "Wikidata")
+                relevant, emb = is_relevant(claim_emb, text, threshold=threshold)
+                if relevant:
+                    save_evidence(text, "Wikidata", embedding=emb)
                     saved += 1
         print(f"[Wikidata] Saved {saved} items.")
         return saved
@@ -273,9 +281,11 @@ def fetch_duckduckgo(claim, claim_emb):
         saved = 0
         for res in results:
             text = res.get_text()
-            if len(text) > 30 and is_relevant(claim_emb, text, 0.05):
-                save_evidence(text, "DuckDuckGo")
-                saved += 1
+            if len(text) > 30:
+                relevant, emb = is_relevant(claim_emb, text, 0.05)
+                if relevant:
+                    save_evidence(text, "DuckDuckGo", embedding=emb)
+                    saved += 1
         print(f"[DuckDuckGo] Saved {saved} items")
     except Exception as e:
         print("[WARNING] DuckDuckGo failed:", e)
@@ -284,15 +294,26 @@ def fetch_duckduckgo(claim, claim_emb):
 # BUILD FAISS
 # ==========================================
 def build_faiss():
+    """Loads pre-calculated embeddings from Database and builds index.
+    No re-encoding performed here — drastically reduces RAM peaks."""
     rows = load_all_evidence()
     if not rows:
         return False
 
-    texts = [row[1] for row in rows]
-    embeddings = embed_model.encode(texts, normalize_embeddings=True)
+    # Filter rows that actually have embeddings
+    texts = []
+    embeddings_list = []
+    for row in rows:
+        if row[3]: # row[3] is the embedding
+            texts.append(row[1])
+            embeddings_list.append(row[3])
+    
+    if not embeddings_list:
+        return False
 
+    embeddings = np.array(embeddings_list).astype('float32')
     index = faiss.IndexFlatIP(embeddings.shape[1])
-    index.add(np.array(embeddings))
+    index.add(embeddings)
 
     faiss.write_index(index, FAISS_FILE)
     return True

project/database.py

@@ -88,20 +88,21 @@ def init_db():
 def clear_db():
     get_db().evidence.delete_many({})
 
-def save_evidence(text, source):
+def save_evidence(text, source, embedding=None):
     try:
         get_db().evidence.insert_one({
             "text":       text,
             "source":     source,
+            "embedding":  embedding, # Store the vector list directly
             "created_at": datetime.now(timezone.utc)
         })
     except Exception as e:
         print(f"[DB] save_evidence error: {e}")
 
 def load_all_evidence():
-    """Returns list of (id, text, source) — same shape the FAISS pipeline expects."""
-    docs = list(get_db().evidence.find({}, {"_id": 1, "text": 1, "source": 1}))
-    return [(str(d["_id"]), d["text"], d["source"]) for d in docs]
+    """Returns list of (id, text, source, embedding) — same shape the FAISS pipeline expects."""
+    docs = list(get_db().evidence.find({}, {"_id": 1, "text": 1, "source": 1, "embedding": 1}))
+    return [(str(d["_id"]), d["text"], d["source"], d.get("embedding")) for d in docs]
 
 def get_total_evidence_count():
     return get_db().evidence.count_documents({})