Deploy SvaraSetu Docker Space

svarasetu/engine.py

@@ -1085,15 +1085,14 @@ class RAGPipelineOrchestrator:
                             stage="gemini_grounded",
                             ms=round((time.perf_counter() - polish_t) * 1000, 2),
                             success=False,
-                            fallback_used=True,
-                            details="Grounded Gemini abstained; trying an uncited open answer",
+                            fallback_used=False,
+                            details="Grounded Gemini abstained; refusing without using general knowledge",
                         ))
-                        return await self._finish_without_corpus(
-                            request=request,
+                        return self._build_declined_response(
                             query=raw_query_text,
                             transcript=transcript,
                             language=target_lang,
-                            reason=gen_details,
+                            reason="The retrieved passages do not contain enough evidence to answer.",
                             guardrails=guardrails,
                             timings=timings,
                             start_t=start_pipeline_t,
@@ -1374,10 +1373,9 @@ class RAGPipelineOrchestrator:
             "UNSAFE_CONTENT",
             "EMPTY_QUERY",
         }
-        gemini_requested = bool(getattr(request, "gemini_enabled", False))
         use_open = (
-            (getattr(config, "USE_GEMINI_OPEN", False) or gemini_requested)
-            and (getattr(config, "ALLOW_NETWORK_CALLS_IN_PIPELINE", False) or gemini_requested)
+            getattr(config, "USE_GEMINI_OPEN", False)
+            and getattr(config, "ALLOW_NETWORK_CALLS_IN_PIPELINE", False)
             and gemini.configured
             and not skip
             and not blocked

svarasetu/types.py

@@ -67,7 +67,7 @@ class QueryRequest(BaseModel):
     cross_lingual: bool = False  # Multilingual federation across all indexed corpora (default: False)
     bypass_cache: bool = False  # Bypass semantic cache to measure cold-start retrieval latency
     bypass_gemini: bool = False  # Force extractive path for the 200 ms SLA bench
-    gemini_enabled: bool = False  # Opt in to grounded polish and open answers on index miss
+    gemini_enabled: bool = False  # Opt in to Gemini synthesis using retrieved evidence only
 
 
 class QueryResponse(BaseModel):

svarasetu/web/index.html

@@ -253,7 +253,7 @@
         <label class="setting-row"><span><strong>Save chat history</strong><small>Keep conversations on this device</small></span><input id="saveHistoryToggle" type="checkbox" checked /><i></i></label>
         <label class="setting-row"><span><strong>Automatic spoken answers</strong><small>Play verified answers after each request</small></span><input id="autoSpeakToggle" type="checkbox" /><i></i></label>
         <label class="setting-row"><span><strong>Cross-lingual retrieval</strong><small>Search across supported languages</small></span><input id="crossLingualToggle" type="checkbox" /><i></i></label>
-        <label class="setting-row"><span><strong>Gemini answers</strong><small>Use Gemini with retrieved evidence, or answer from general knowledge when the index has no match</small></span><input id="geminiToggle" type="checkbox" /><i></i></label>
+        <label class="setting-row"><span><strong>Gemini + RAG</strong><small>Let Gemini answer only from passages retrieved by your RAG index</small></span><input id="geminiToggle" type="checkbox" /><i></i></label>
       </section>
       <section class="settings-group">
         <h3>Appearance</h3>

tests/test_api.py

@@ -108,6 +108,8 @@ def test_home_ui_and_styles(client):
     assert "ingestBtn" not in home.text
     assert "Evidence-backed answers across 15 indexed languages" not in home.text
     assert 'id="geminiToggle"' in home.text
+    assert "answer from general knowledge" not in home.text
+    assert "Gemini + RAG" in home.text
     assert "lang-btn" not in home.text
     assert "Tap once. Then talk naturally." in home.text
     assert "SILENCE_STOP_MS" in home.text

tests/test_gemini_quality.py

@@ -1,11 +1,15 @@
 """Gemini quality layer: embed rerank + live client. No network."""
 
+import asyncio
 import sys
+import time
 from pathlib import Path
 
 sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
 
 from svarasetu.compose.polish import GeminiClient, _cosine, get_gemini_client
+from svarasetu.engine import RAGPipelineOrchestrator
+from svarasetu.types import GuardrailFlags, QueryRequest
 
 
 class TestCosine:
@@ -85,6 +89,32 @@ class TestGeminiOpen:
         )
         assert resp.answer_source == "gemini_open"
 
+    def test_request_gemini_setting_never_enables_open_answers(self, monkeypatch):
+        class FakeGemini:
+            configured = True
+
+            def generate_open(self, *_args, **_kwargs):
+                raise AssertionError("open-domain Gemini must not run for the RAG toggle")
+
+        monkeypatch.setattr("svarasetu.engine.get_gemini_client", lambda: FakeGemini())
+        monkeypatch.setattr("svarasetu.engine.config.USE_GEMINI_OPEN", False)
+        monkeypatch.setattr("svarasetu.engine.config.ALLOW_NETWORK_CALLS_IN_PIPELINE", False)
+
+        pipeline = RAGPipelineOrchestrator.__new__(RAGPipelineOrchestrator)
+        response = asyncio.run(
+            pipeline._finish_without_corpus(
+                request=QueryRequest(text="Not in the index", gemini_enabled=True),
+                query="Not in the index",
+                transcript="Not in the index",
+                language="en",
+                reason="No indexed evidence",
+                guardrails=GuardrailFlags(decline_reason_code="LOW_RETRIEVAL_CONFIDENCE"),
+                timings=[],
+                start_t=time.perf_counter(),
+            )
+        )
+        assert response.answer_source == "declined"
+
 
 class TestGeminiClientFactory:
     def test_unconfigured_without_key(self):