Upload folder using huggingface_hub

app.py

@@ -359,6 +359,24 @@ chatInterface = gr.ChatInterface(
     additional_inputs=[thread_id, debug_checkbox],
     chatbot=chatbot,
     css="""
+table {
+  border-collapse: collapse;
+  width: 90%;
+}
+
+table, th, td {
+  border: 1px solid #ddd;
+  padding: 6px;
+  font-size: small;
+}
+
+td {
+  word-wrap: break-word;
+  white-space: pre-wrap; /* preserves line breaks but wraps long lines */
+  max-width: 300px; /* control width */
+  vertical-align: top;
+}
+
 .spinner {
   display: inline-block;
   width: 1em;

graph_helper.py

@@ -73,68 +73,79 @@ def generate_graph() -> CompiledStateGraph:
             SystemMessage(
                 content=(
                     """
-                        You are a response validator. Your job is to:
-
-                        1. Fix encoding issues in native-language verses (e.g., Tamil, Sanskrit) **only** for garbled/malformed characters. Do NOT paraphrase, invent, or alter meaning.
-
-                        2. Validate content authenticity:
-                        - Extract key entities from the original user query and from the assistant's response. Key entities include pasuram/verse numbers, azhwar names, prabandham names, divya desam names, deity/context (e.g., "Krishna playing flute"), etc.
-                        - Compare them literally (exact matches for numeric IDs such as pasuram numbers). If a required entity (e.g., pasuram number) does not match exactly, mark it INCORRECT.
-                        - List any missing, altered, or hallucinated entities.
-
-                        3. Return the corrected assistant message **first**. After the corrected message, append a standalone HTML evaluation block that begins on its own line (i.e., include a blank line before the block). **Do not** wrap the evaluation block in Markdown fences. The evaluation block MUST be either:
-
-                        A. An HTML block using a proper HTML `<table>` (preferred). Use the following structure and styles:
-
-                        <div style="font-size:0.9em; opacity:0.5; margin-top:10px;">
-                            <hr>
-                            <strong>Original user query:</strong> ...<br/><br/>
-                            <strong>Entity comparison:</strong><br/>
-                            <table style="border-collapse:collapse; margin-top:6px;">
-                            <thead>
-                                <tr>
-                                <th style="border:1px solid #ddd; padding:6px; text-align:left;">Expected</th>
-                                <th style="border:1px solid #ddd; padding:6px; text-align:left;">Found</th>
-                                </tr>
-                            </thead>
-                            <tbody>
-                                <tr><td style="border:1px solid #ddd; padding:6px;">value</td><td style="border:1px solid #ddd; padding:6px;">value</td></tr>
-                            </tbody>
-                            </table>
-                            <br/>
-                            <strong>Verdict:</strong> ...<br/>
-                            <strong>Confidence:</strong> ...<br/>
-                            <strong>Badge:</strong> <span style="display:inline-block; padding:3px 8px; border-radius:6px; color:#fff; font-weight:600;">...</span><br/>
-                            <strong>Note:</strong> ...
-                        </div>
-
-                        B. **If** the environment strips `<table>` tags, fall back to an ASCII table inside a `<pre>` within the styled `<div>` (monospace + same opacity). Example fallback:
-
-                        <div style="font-size:0.9em; opacity:0.75; margin-top:10px; font-family:monospace;">
-                        <pre>
-                        Original user query: ...
-                        
-                        Entity comparison:
-                        
-                        | Expected | Found |
-                        |----------|-------|
-                        | 1145     | 696   |
-                        
-                        Verdict: Incorrect
-                        Confidence: 40% — The verse number does not match.
-                        Badge: 🔴
-                        Note: ...
-                        </pre>
-                        </div>
-
-                        4. Always leave a blank line before the HTML `<div>` so renderers treat it as a separate block.
-
-                        5. Be strict: if the user asked for pasuram 1145 but the assistant returned 696 (or any other number), mark the verdict **Incorrect** and set confidence under 50% (e.g., 30–45%) with a short justification.
-
-                        6. If you fixed native text, explicitly list the fixes in the Note.
-
-                        Return only: (a) the corrected assistant message, and then (b) the standalone evaluation HTML block described above (no extra explanation).
-                    
+You are a strict validator for LLM responses to scripture queries.
+
+Your tasks:
+0. Treat your input as `original_llm_response`.
+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.
+4. Verify (from `original_llm_response`):
+- 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:
+---
+<!-- **Step 1 – Repaired LLM Response in Markdown:** -->
+<!-- BEGIN_MARKDOWN -->
+fixed_llm_response
+<!-- END_MARKDOWN -->
+
+<!-- **Step 2 – Validation Table:** -->
+<div style="font-size: small; opacity: 0.6;">
+<hr>
+<b>Original user query:</b> {{original_user_query}}
+
+<table border="1" cellpadding="4" cellspacing="0" style="border-collapse: collapse; width: 100%;">
+<tr>
+<th>Expected</th>
+<th>Found</th>
+<th>Match?</th>
+</tr>
+<tr>
+<td>verse number(s)</td>
+<td>{{found_verse_numbers}}</td>
+<td>{{match_status_for_verse}}</td>
+</tr>
+<tr>
+<td>keyword/context</td>
+<td>{{found_keyword}}</td>
+<td>{{match_status_for_keyword}}</td>
+</tr>
+{{dynamic_entity_rows}}
+<tr>
+<td>native verse text</td>
+<td style="white-space: normal; word-break: break-word;">{{cleaned_native_text}}</td>
+<td>{{garbled_fix_status}}</td>
+</tr>
+</table>
+
+<p><b>Verdict:</b> {{Verdict}}<br>
+<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`.
+- ✅, ❌, ⚠️ remain for matches.
+- Hidden markers (`<!-- BEGIN_MARKDOWN -->`) make it parseable without showing literal marker text.
+
+
                     """
                 )
             ),

sanatan_assistant.py

@@ -94,14 +94,14 @@ Respond in **Markdown** format only. Ensure Sanskrit/Tamil verses are always cle
     return prompt
 
 
-def query(collection_name: allowedCollections, query: str, n_results=5):
+def query(collection_name: allowedCollections, query: str, n_results=3):
     """
     Search a scripture collection.
 
     Parameters:
     - collection_name (str): The name of the scripture collection to search. ...
     - query (str): The search query.
-    - n_results (int): Number of results to return. Default is 5.
+    - n_results (int): Number of results to return. Default is 3.
 
     Returns:
     - A list of matching results.
@@ -126,7 +126,7 @@ def query_by_metadata_field(
         "$in", "$eq", "$gt", "$gte", "$lt", "$lte", "$ne", "$nin"
     ],
     metadata_value: Any,
-    n_results=5,
+    n_results=3,
 ):
     """
     Search a scripture collection by metadata. Do NOT use this for semantic search. Only use when a specific metadata field is provided.
@@ -137,7 +137,7 @@ def query_by_metadata_field(
     - metadata_field (str) : The name of the metadata field. e.g. azhwar_name
     - metadata_search_operator (str) : The search operator e.g. $eq or $in. DO NOT use $regex.
     - metadata_value : Value to search for can be any primitive datatype like str or int (or a list[str] if metadata_search_operator = '$in'). for e.g. Thirumangai Azhwar or '2233' or 2233
-    - n_results (int): Number of results to return. Default is 5.
+    - n_results (int): Number of results to return. Default is 3.
 
     Returns:
     - A list of matching results.
@@ -169,7 +169,7 @@ def query_by_metadata_field(
 def query_by_literal_text(
     collection_name: allowedCollections,
     literal_to_search_for: str,
-    n_results=5,
+    n_results=3,
 ):
     """
     Search a scripture collection by a literal. Do NOT use this for semantic search. Only use when the user specifically asks for literal search.
@@ -177,7 +177,7 @@ def query_by_literal_text(
     Parameters:
     - collection_name (str): The name of the scripture collection to search. ...
     - literal_to_search_for (str): The search query.
-    - n_results (int): Number of results to return. Default is 5.
+    - n_results (int): Number of results to return. Default is 3.
 
     Returns:
     - A list of matching results.