Upload folder using huggingface_hub

app.py

@@ -295,6 +295,18 @@ async def chat_streaming(message, history, thread_id):
                     logger.debug(
                         f"Partial arguments for tool {accumulated_tool_call['name']}: {accumulated_tool_call['args_str']}"
                     )
+    except asyncio.CancelledError:
+        logger.warning("⚠️ Request cancelled by user")
+        node_tree = end_node_tree(node_tree=node_tree)
+        yield (
+            f"### {' → '.join(node_tree)}"
+            "\n⚠️⚠️⚠️ Request cancelled by user"
+            "\nhere is what I got so far ...\n"
+            f"\n{streamed_response}"
+        )
+        # Important: re-raise if you want upstream to also know
+        # raise      
+        return                
     except Exception as e:
         logger.error("❌❌❌ Error processing request: %s", e)
         traceback.print_exc()
@@ -305,7 +317,7 @@ async def chat_streaming(message, history, thread_id):
             "\nhere is what I got so far ...\n"
             f"\n{streamed_response}"
         )
-
+        return
 
 # UI Elements
 thread_id = gr.State(init_session)

graph_helper.py

@@ -80,23 +80,36 @@ Your tasks:
 1. Compare the original user query to the LLM’s answer.
 2. Identify the scripture context (e.g., Divya Prabandham, Bhagavad Gita, Upanishads, Ramayana, etc.).
 3. Based on the scripture context, dynamically choose the appropriate entity columns for validation:
-- **Divya Prabandham** → azhwar, prabandham, location/deity
-- **Bhagavad Gita** → chapter, verse number(s), speaker, listener
-- **Upanishads** → section, mantra number, rishi, deity
-- **Ramayana/Mahabharata** → book/kanda, section/sarga, character(s), location
-- **Other** → pick the 3–4 most relevant contextual entities from the scripture’s metadata.
+   - **Divya Prabandham** → azhwar, prabandham, location/deity
+   - **Bhagavad Gita** → chapter, verse number(s), speaker, listener
+   - **Upanishads** → section, mantra number, rishi, deity
+   - **Ramayana/Mahabharata** → book/kanda, section/sarga, character(s), location
+   - **Other** → pick the 3–4 most relevant contextual entities from the scripture’s metadata.
 4. Verify (from `original_llm_response`):
-- Correct verse number(s)
-- Keyword/context match
-- All scripture-specific entity fields
-- Native verse text quality
+   - Correct verse number(s)
+   - Keyword/context match
+   - All scripture-specific entity fields
+   - Native verse text quality
 5. **Repair any garbled Tamil/Sanskrit characters** in the verse:
-- Restore correct letters, diacritics, and punctuation.
-- Replace broken Unicode with proper characters.
-- Correct vowel signs, consonants, and pulli markers.
-- Preserve original spacing and line breaks.
-The repaired version is `fixed_llm_response`.
-6. Output in this exact order:
+   - Restore correct letters, diacritics, and punctuation.
+   - Replace broken Unicode with proper characters.
+   - Correct vowel signs, consonants, and pulli markers.
+   - Preserve original spacing and line breaks.
+   The repaired version is `fixed_llm_response`.
+
+6. Confidence-based display rule:
+   - If `Confidence` ≥ 75:
+       1. Show **Step 1** and **Step 2** as described below.
+   - If `Confidence` < 75:
+       - Show only:
+         ```
+         I got a response for your query, but I am not confident enough in its accuracy to display it in full.
+         Confidence score: {{Confidence}}%
+         ```
+       - Do not include the repaired LLM response or the validation table in this case.
+
+7. If showing the full output (Confidence ≥ 75), follow this format exactly:
+
 ---
 <!-- **Step 1 – Repaired LLM Response in Markdown:** -->
 <!-- BEGIN_MARKDOWN -->
@@ -136,15 +149,14 @@ fixed_llm_response
 <b>Confidence score:</b> {{Confidence}}% – {{Justification}}<br>
 <span style="background-color:{{badge_color_code}}; color:white; padding:2px 6px; border-radius:4px;">{{badge_emoji}}</span></p>
 </div>
-
 ---
 
 Where:
 - `{{dynamic_entity_rows}}` is context-specific entity rows.
-- `{{cleaned_native_text}}` must be taken from the repaired `fixed_llm_response`.
+- `{{cleaned_native_text}}` must be from the repaired `fixed_llm_response`.
 - ✅, ❌, ⚠️ remain for matches.
-- Hidden markers (`<!-- BEGIN_MARKDOWN -->`) make it parseable without showing literal marker text.
-
+- Hidden markers (`<!-- BEGIN_MARKDOWN -->`) prevent them from rendering as visible text.
+- Always wrap verse text so it doesn’t overflow horizontally.
 
                     """
                 )