fix

src/streamlit_app.py

@@ -20,7 +20,7 @@ from typing import Dict, List, Optional, Tuple
 
 import requests
 import streamlit as st
-from llama_index.core import SimpleDirectoryReader, SummaryIndex, VectorStoreIndex
+from llama_index.core import Settings, SimpleDirectoryReader, SummaryIndex, VectorStoreIndex
 from llama_index.core.agent.workflow import FunctionAgent
 from llama_index.core.node_parser import SentenceSplitter
 from llama_index.core.objects import ObjectIndex
@@ -48,13 +48,21 @@ DOC_SOURCES: Dict[str, Tuple[str, str, str]] = {
     ),
 }
 BASE_SYSTEM_PROMPT = (
-    "You are an agent designed to answer queries over a set of RYBREVANT documents." \
-    "Please always use the tools provided to answer a question. Do not rely on prior knowledge." \
-    "When responding, keep answers concise, always mention the source: exact document + page "
-    "(e.g., 'PI p.12' or 'brochure p.5'), and end with a brief safety disclaimer "
-    "('Not medical advice; consult your healthcare professional')."
+    "You are an agent designed to answer queries over a set of RYBREVANT documents. "
+    "Always use the tools provided to answer a question. Do not rely on prior knowledge or any information "
+    "outside the provided documents. "
+    "When responding, keep answers concise, always mention the source with exact document + page "
+    "(for example, 'PI p.12' or 'brochure p.5'), and end with a brief safety disclaimer "
+    "('Not medical advice; consult your healthcare professional'). "
+    "If a question cannot be fully answered from the documents, clearly say so and recommend consulting a "
+    "healthcare professional rather than guessing or extrapolating. "
+    "For detailed, safety-critical questions such as dosing, contraindications, warnings, or administration "
+    "details, always prefer the vector tools over summary tools so that answers are grounded in the exact text."
 )
 
+# Global embedding model used across all indices
+Settings.embed_model = OpenAIEmbedding(model="text-embedding-3-large")
+
 
 def _ensure_data_files() -> None:
     """Download source PDFs if they are missing."""
@@ -104,8 +112,7 @@ def _build_tools_for_doc(file_path: Path, name: str):
     if not nodes:
         raise ValueError(f"No text nodes parsed from {file_path}. Check parser credentials or PDF availability.")
 
-    embed_model = OpenAIEmbedding(model="text-embedding-3-large")
-    vector_index = VectorStoreIndex(nodes, embed_model=embed_model)
+    vector_index = VectorStoreIndex(nodes)
 
     def vector_query(query: str, page_numbers: Optional[List[int]] = None) -> str:
         """Grounded Q&A with optional page filters + citations.
@@ -215,6 +222,9 @@ def main() -> None:
             "- Data/parsing cached in this Space runtime"
         )
         st.divider()
+        if st.button("Clear conversation", type="secondary"):
+            st.session_state.pop("messages", None)
+            st.experimental_rerun()
         st.markdown("Need to deploy? Push this app to Hugging Face Spaces with your API keys as secrets.")
 
     has_keys = _require_env("OPENAI_API_KEY") and _require_env("LLAMA_CLOUD_API_KEY")