1.0

.env

@@ -0,0 +1,2 @@
+GOOGLE_API_KEY=AIzaSyBapGOjPJR58TlTMYcnz7G1jP8fsJXZ1Tg
+NV_API_KEY=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

.streamlit/config.toml

@@ -0,0 +1,5 @@
+[server]
+headless = true
+
+[browser]
+gatherUsageStats = false

Dockerfile

@@ -8,4 +8,4 @@ RUN apt-get update && apt-get install -y graphviz
 
 COPY . .
 
-CMD ["streamlit", "run", "app.py", "--server.port", "7860", "--server.address", "0.0.0.0"]
+CMD ["streamlit", "run", "app.py", "--server.port", "7860", "--server.address", "0.0.0.0","--server.enableCORS=false", "--server.enableXsrfProtection=false"]

app.py

@@ -74,6 +74,8 @@ if "deep_summary" not in st.session_state:
     st.session_state.deep_summary = None
 if "graph_dot" not in st.session_state:
     st.session_state.graph_dot = None
+if "podcast_audio" not in st.session_state:
+    st.session_state.podcast_audio = None
 
 def switch_page(page_name):
     st.session_state.page = page_name
@@ -244,7 +246,7 @@ def view_summary_dialog(text):
 
 @dialog_decorator("Knowledge Graph Visualization", width="large")
 def view_graph_dialog(dot_code):
-    st.graphviz_chart(dot_code, width='stretch')
+    st.graphviz_chart(dot_code, width="stretch")
     st.caption("Right-click -> 'Open Image in New Tab' to zoom/download.")
 
 def show_app():
@@ -288,6 +290,7 @@ def show_app():
                 st.session_state.full_text = ""
                 st.session_state.processed_files = set()
                 st.session_state.upload_status = ""
+                st.session_state.podcast_audio = None
                 st.session_state.uploader_key += 1
                 st.rerun()
 
@@ -301,6 +304,10 @@ def show_app():
         # Chat History
         for msg in st.session_state.messages:
             with st.chat_message(msg["role"]):
+                if "thoughts" in msg and msg["thoughts"]:
+                    with st.expander("⛓️ Reasoning Log", expanded=False):
+                        for log in msg["thoughts"]:
+                            st.write(log)
                 st.markdown(msg["content"])
         
         # User Input
@@ -313,22 +320,26 @@ def show_app():
                 if st.session_state.agent:
                     # Container for intermediate thought process
                     with st.status("Agent Reasoning...", expanded=True) as status:
+                        thoughts = []
                         
                         def graph_callback(node_name, state):
-                            if node_name == "expand_query":
-                                status.write(f"🧠 **Expanding Query** with related concepts...")
-                            elif node_name == "retrieve":
-                                status.write(f"🔍 **Retrieving** context for query: *'{state.get('current_query', '...')}'*")
+                            msg = ""
+                            if node_name == "retrieve":
+                                msg = f"🔍 **Retrieving** context for query: *'{state.get('current_query', '...')}'*"
                             elif node_name == "generate":
-                                status.write("🧠 **Generating** answer...")
+                                msg = "🧠 **Generating** answer..."
                             elif node_name == "reflect":
                                 score = state.get("reflection_score")
                                 if score == "yes":
-                                    status.write("✅ **Reflection Passed**: Answer is grounded.")
+                                    msg = "✅ **Reflection Passed**: Answer is grounded."
                                 else:
-                                    status.write("❌ **Reflection Failed**: Hallucination/Irrelevance detected.")
+                                    msg = "❌ **Reflection Failed**: Hallucination/Irrelevance detected."
                             elif node_name == "rewrite_query":
-                                status.write(f"🔄 **Rewriting Query** to improve results...")
+                                msg = f"🔄 **Rewriting Query** to improve results..."
+                            
+                            if msg:
+                                status.write(msg)
+                                thoughts.append(msg)
 
                         result = st.session_state.agent.run(prompt, callback=graph_callback)
                         status.update(label="Response Ready", state="complete", expanded=False)
@@ -336,11 +347,15 @@ def show_app():
                         response = result["generation"]
                         
                         # Show debug steps comfortably (Optional redundant info, maybe keep for final stats)
-                        with st.expander("⛓️ Final Stats", expanded=False):
+                        with st.expander("📊 Final Stats", expanded=False):
                             st.write(f"**Reflected:** {result.get('reflection_score')} | **Total Iter:** {result.get('iterations')}")
                         
                         st.markdown(response)
-                        st.session_state.messages.append({"role": "assistant", "content": response})
+                        st.session_state.messages.append({
+                            "role": "assistant", 
+                            "content": response,
+                            "thoughts": thoughts
+                        })
                 else:
                     st.warning("Please upload a PDF first.")
 
@@ -365,22 +380,24 @@ def show_app():
             
             # Podcast Tool
             with st.expander("🎧 Podcast", expanded=False):
-                if st.button("Generate Audio"):
-                    briefing = ensure_deep_summary()
-                    
-                    with st.spinner("Scripting & Synthesizing..."):
-                        p_gen = PodcastGenerator()
-                        script = p_gen.generate_audio_script(briefing)
-                        
-                        # Show script preview
-                        st.caption("Dialogue generated.")
-                        
-                        # Generate Audio
-                        audio_path = p_gen.generate_audio_file(script)
-                        if audio_path:
-                            st.audio(audio_path)
-                        else:
-                            st.error("Audio generation failed.")
+                if not st.session_state.podcast_audio:
+                    if st.button("Generate Audio"):
+                        briefing = ensure_deep_summary()
+                        with st.spinner("Scripting & Synthesizing..."):
+                            p_gen = PodcastGenerator()
+                            script = p_gen.generate_audio_script(briefing)
+                            audio_path = p_gen.generate_audio_file(script)
+                            if audio_path:
+                                st.session_state.podcast_audio = audio_path
+                                st.rerun()
+                            else:
+                                st.error("Audio generation failed.")
+                else:
+                    st.success("Podcast Ready!")
+                    st.audio(st.session_state.podcast_audio)
+                    if st.button("🔄 Regenerate Podcast"):
+                         st.session_state.podcast_audio = None
+                         st.rerun()
 
             # Knowledge Graph Tool
             with st.expander("🕸️ Knowledge Graph", expanded=False):

core/graph.py

@@ -2,7 +2,7 @@ from typing import TypedDict, List
 from langgraph.graph import StateGraph, END
 from langchain_core.documents import Document
 from core.models import get_llm
-from prompts import RAG_PROMPT, REFLECTION_PROMPT, REWRITE_PROMPT, QUERY_EXPANSION_PROMPT
+from prompts import RAG_PROMPT, REFLECTION_PROMPT, REWRITE_PROMPT
 from langchain_core.output_parsers import StrOutputParser
 
 class GraphState(TypedDict):
@@ -19,12 +19,6 @@ class RAGAgent:
         self.llm = get_llm()
         self.app = self.build_graph()
 
-    def expand_query(self, state: GraphState):
-        question = state["question"]
-        chain = QUERY_EXPANSION_PROMPT | self.llm | StrOutputParser()
-        expanded_query = chain.invoke({"question": question})
-        return {"current_query": expanded_query}
-
     def retrieve(self, state: GraphState):
         query = state["current_query"]
         docs = self.retriever.invoke(query)
@@ -44,9 +38,17 @@ class RAGAgent:
     def reflect(self, state: GraphState):
         question = state["question"]
         generation = state["generation"]
+        docs = state["documents"]
+        
+        # Format context so the reflector can check for grounding
+        context = "\n\n".join([f"[Source: {doc.metadata.get('filename', 'Unknown')}] {doc.page_content}" for doc in docs])
         
         chain = REFLECTION_PROMPT | self.llm | StrOutputParser()
-        score = chain.invoke({"question": question, "generation": generation})
+        score = chain.invoke({
+            "context": context,
+            "question": question, 
+            "generation": generation
+        })
         
         # Normalize score
         normalized_score = "yes" if "yes" in score.lower() else "no"
@@ -54,9 +56,15 @@ class RAGAgent:
 
     def rewrite_query(self, state: GraphState):
         question = state["question"]
+        previous_query = state["current_query"]
+        failed_gen = state["generation"]
         
         chain = REWRITE_PROMPT | self.llm | StrOutputParser()
-        new_query = chain.invoke({"question": question})
+        new_query = chain.invoke({
+            "question": question, 
+            "previous_query": previous_query, 
+            "generation": failed_gen
+        })
         
         return {"current_query": new_query, "iterations": state["iterations"] + 1}
 
@@ -73,15 +81,13 @@ class RAGAgent:
         workflow = StateGraph(GraphState)
 
         # Define Nodes
-        workflow.add_node("expand_query", self.expand_query)
         workflow.add_node("retrieve", self.retrieve)
         workflow.add_node("generate", self.generate)
         workflow.add_node("reflect", self.reflect)
         workflow.add_node("rewrite_query", self.rewrite_query)
 
         # Build Edges
-        workflow.set_entry_point("expand_query")
-        workflow.add_edge("expand_query", "retrieve")
+        workflow.set_entry_point("retrieve")
         workflow.add_edge("retrieve", "generate")
         workflow.add_edge("generate", "reflect")
         

core/models.py

@@ -1,7 +1,8 @@
 import os
 import streamlit as st
 from langchain_nvidia_ai_endpoints import ChatNVIDIA
-from langchain_google_genai import GoogleGenerativeAIEmbeddings
+from langchain_google_genai import GoogleGenerativeAIEmbeddings, ChatGoogleGenerativeAI
+from google import genai
 
 def get_llm(model_name: str = "nvidia/nemotron-3-nano-30b-a3b"):
     """
@@ -37,3 +38,51 @@ def get_embeddings():
         raise ValueError("GOOGLE_API_KEY not found in environment or secrets.")
         
     return GoogleGenerativeAIEmbeddings(model="models/embedding-001", google_api_key=api_key)
+
+from google.genai import types
+
+
+def generate_podcast_audio(script_text: str):
+    """
+    Calls Gemini TTS with multi-speaker configuration.
+    Returns raw audio data.
+    """
+    api_key = os.getenv("GOOGLE_API_KEY")
+    if not api_key and "GOOGLE_API_KEY" in st.secrets:
+        api_key = st.secrets["GOOGLE_API_KEY"]
+    
+    client = genai.Client(api_key=api_key)
+    
+    response = client.models.generate_content(
+        model="gemini-2.5-flash-preview-tts",
+        contents=f"Generate a podcast dialogue audio. \n\n{script_text}",
+        config=types.GenerateContentConfig(
+            response_modalities=["AUDIO"],
+            speech_config=types.SpeechConfig(
+                multi_speaker_voice_config=types.MultiSpeakerVoiceConfig(
+                    speaker_voice_configs=[
+                        types.SpeakerVoiceConfig(
+                            speaker='Alex',
+                            voice_config=types.VoiceConfig(
+                                prebuilt_voice_config=types.PrebuiltVoiceConfig(
+                                    voice_name='Kore',
+                                )
+                            )
+                        ),
+                        types.SpeakerVoiceConfig(
+                            speaker='Jamie',
+                            voice_config=types.VoiceConfig(
+                                prebuilt_voice_config=types.PrebuiltVoiceConfig(
+                                    voice_name='Puck',
+                                )
+                            )
+                        ),
+                    ]
+                )
+            )
+        )
+    )
+    
+    if response.candidates and response.candidates[0].content.parts:
+        return response.candidates[0].content.parts[0].inline_data.data
+    return None

core/pdf_processer.py

@@ -76,7 +76,6 @@ class PDFProcessor:
     def get_retriever(self):
         if not self.vector_store:
             raise ValueError("Vector store not initialized. Upload a PDF first.")
-        # return self.vector_store.as_retriever(search_type="mmr", search_kwargs={"k": 5})
         return self.vector_store.as_retriever(search_type="mmr", search_kwargs={"k": 5})
     def get_full_text(self):
         return "\n\n".join([doc.page_content for doc in self.documents])

core/podcast.py

@@ -1,13 +1,9 @@
-from core.models import get_llm
+from core.models import get_llm, generate_podcast_audio
 from prompts import PODCAST_AUDIO_PROMPT
 from langchain_core.prompts import ChatPromptTemplate
 from langchain_core.output_parsers import StrOutputParser
 import tempfile
-import os
-from google import genai
-from google.genai import types
 import wave
-import streamlit as st
 
 class PodcastGenerator:
     def __init__(self):
@@ -25,52 +21,12 @@ class PodcastGenerator:
 
     def generate_audio_file(self, script_text):
         """
-        Uses 'gemini-2.5-flash-preview-tts' with official google-genai SDK for multi-speaker.
+        Uses centralized Gemini TTS logic.
         """
-        api_key = os.getenv("GOOGLE_API_KEY")
-        if not api_key and "GOOGLE_API_KEY" in st.secrets:
-            api_key = st.secrets["GOOGLE_API_KEY"]
-            
-        if not api_key:
-            return None
-
-        client = genai.Client(api_key=api_key)
-
         try:
-            response = client.models.generate_content(
-                model="gemini-2.5-flash-preview-tts",
-                contents=f"Generate a podcast dialogue audio. \n\n{script_text}",
-                config=types.GenerateContentConfig(
-                    response_modalities=["AUDIO"],
-                    speech_config=types.SpeechConfig(
-                        multi_speaker_voice_config=types.MultiSpeakerVoiceConfig(
-                            speaker_voice_configs=[
-                                types.SpeakerVoiceConfig(
-                                    speaker='Alex',
-                                    voice_config=types.VoiceConfig(
-                                        prebuilt_voice_config=types.PrebuiltVoiceConfig(
-                                            voice_name='Kore',
-                                        )
-                                    )
-                                ),
-                                types.SpeakerVoiceConfig(
-                                    speaker='Jamie',
-                                    voice_config=types.VoiceConfig(
-                                        prebuilt_voice_config=types.PrebuiltVoiceConfig(
-                                            voice_name='Puck',
-                                        )
-                                    )
-                                ),
-                            ]
-                        )
-                    )
-                )
-            )
-
-            # Extract audio data
-            if response.candidates and response.candidates[0].content.parts:
-                data = response.candidates[0].content.parts[0].inline_data.data
-                
+            data = generate_podcast_audio(script_text)
+            
+            if data:
                 # Use NamedTemporaryFile to get a unique name, then close it immediately
                 # so wave.open can re-open it for writing.
                 with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp:

core/visualizer.py

@@ -3,7 +3,7 @@ from langchain_core.prompts import PromptTemplate
 from langchain_core.output_parsers import StrOutputParser
 import graphviz
 import re
-from prompts import GRAPH_TEMPLATE
+from prompts import GRAPH_PROMPT
 
 class KnowledgeGraphGenerator:
     def __init__(self):
@@ -13,7 +13,7 @@ class KnowledgeGraphGenerator:
         # Text is now the "Deep Summary", so no need to truncate.
         input_text = text 
         
-        chain = PromptTemplate.from_template(GRAPH_TEMPLATE) | self.llm | StrOutputParser()
+        chain = GRAPH_PROMPT | self.llm | StrOutputParser()
         dot_code = chain.invoke({"text": input_text})
         
         # Cleanup markdown if present

prompts.py

@@ -1,13 +1,18 @@
 from langchain_core.prompts import ChatPromptTemplate
 
 # RAG Generation Prompt
-RAG_SYSTEM = """You are an academic research assistant. Answer the user's question based strictly on the provided context. 
+RAG_SYSTEM = """You are a research assistant. Answer the user's question based strictly on the provided context. 
 If the context does not contain the answer, say "I cannot answer this based on the document."
 
 Requirements:
 1. Use academic tone.
-2. YOU MUST CITE sources using (Document: <filename>, Page: <page>) format at the end of relevant sentences.
-3. Be concise but comprehensive."""
+2. **In-text Citations:** Use Unicode Superscript Numbers (¹, ², ³, ⁴, ⁵, ⁶, ⁷, ⁸, ⁹, ¹⁰) strictly. Place them immediately after the punctuation or relevant phrase.
+   - Do NOT use `[^1]` (Markdown footnotes) or `[1]` (Brackets). 
+   - Example: ...at compile time¹.
+3. **References Section:** At the very end, include a section titled "References".
+4. **Reference Format:** List the citations sequentially using normal numbers.
+   - Format: `1. Document: <filename>, Page: <page>`
+5. Be concise but comprehensive."""
 
 RAG_HUMAN = """Context:
 {context}
@@ -24,13 +29,17 @@ RAG_PROMPT = ChatPromptTemplate.from_messages([
 
 # Reflection Prompt
 REFLECTION_SYSTEM = """You are a senior editor grading an AI-generated answer. 
-Check if the answer is grounded in the provided documents and relevant to the user's question.
+Evaluate if the answer is:
+1. GROUNDED: Is the answer supported by the facts in the provided Context? 
+2. RELEVANT: Does it actually answer the User Question?
 
-Output exactly "yes" if the answer is grounded and relevant.
-Output "no" if the answer is hallucinated, irrelevant, or incomplete.
-IMPORTANT: If the answer says "I cannot answer" or "context does not contain", YOU MUST OUTPUT "no"."""
+Output exactly "yes" if the answer is both grounded and relevant.
+Output "no" if the answer contains information NOT in the context, is irrelevant, or if the assistant says it cannot answer."""
 
-REFLECTION_HUMAN = """User Question: {question}
+REFLECTION_HUMAN = """Context:
+{context}
+
+User Question: {question}
 Generated Answer: {generation}
 
 Current Answer Quality status:"""
@@ -41,59 +50,61 @@ REFLECTION_PROMPT = ChatPromptTemplate.from_messages([
 ])
 
 # Query Rewrite Prompt
-REWRITE_SYSTEM = """You are a query optimizer. The previous search query failed to retrieve relevant documents.
-Rewrite the user's question to be more specific and keyword-rich for vector retrieval.
+REWRITE_SYSTEM = """You are a query optimizer. The previous search query failed to retrieve documents that could fully answer the question.
+Analyze the original question, the previous query used, and the failed answer to understand what was missing or misunderstood.
+Rewrite the pursuit into a new, improved search query that is more specific and uses better technical keywords.
 Output ONLY the rewritten query string."""
 
 REWRITE_HUMAN = """Original Question: {question}
+Previous Query: {previous_query}
+Failed Answer: {generation}
 
-Rewritten Query:"""
+Improved Rewritten Query:"""
 
 REWRITE_PROMPT = ChatPromptTemplate.from_messages([
     ("system", REWRITE_SYSTEM),
     ("human", REWRITE_HUMAN)
 ])
 
-# Query Expansion Prompt (Pre-retrieval)
-QUERY_EXPANSION_SYSTEM = """You are a research assistant. 
-Rewrite the user's query to be more effective for vector retrieval (RAG).
-- Add 2-3 relevant academic keywords, synonyms, or related concepts.
-- Keep the original intent and core subject intact.
-- Output ONLY the rewritten/expanded query string. No explanations."""
-
-QUERY_EXPANSION_HUMAN = "Original Query: {question}\n\nExpanded Query:"
-
-QUERY_EXPANSION_PROMPT = ChatPromptTemplate.from_messages([
-    ("system", QUERY_EXPANSION_SYSTEM),
-    ("human", QUERY_EXPANSION_HUMAN)
-])
-
 # Podcast Prompts
 # Summary/Podcast Map Prompts
-SUMMARY_MAP_SYSTEM = """Summarize the following chunk of text for a briefing. 
-Limit your response to a maximum of 500 words. Focus on key facts, methodology, and results."""
+SUMMARY_MAP_SYSTEM = """You are a precision-oriented research analyst. 
+Extract atomic facts and technical details from this segment into dense bullet points.
+
+STRICT CONSTRAINT: Maximum 500 words total.
+Requirements:
+1. Highlight core concepts and key relationships (e.g., "A influences B").
+2. Maintain technical accuracy and preserve specialized terminology.
+3. Be concise: avoid introductory phrases, focus on pure data/logic."""
 
 SUMMARY_MAP_HUMAN = """Text Chunk:
 {text}
 
-Summary:"""
+Atomic Fact Summary:"""
 
 SUMMARY_MAP_PROMPT = ChatPromptTemplate.from_messages([
     ("system", SUMMARY_MAP_SYSTEM),
     ("human", SUMMARY_MAP_HUMAN)
 ])
 
-SUMMARY_REDUCE_SYSTEM = """Synthesize the following summaries into a structured "Deep Briefing".
-The total length must not exceed 2000 words.
-The briefing should have:
-1. Main Theme
-2. Key Findings
-3. Methodology
-4. Implications"""
+SUMMARY_REDUCE_SYSTEM = """You are a Senior Knowledge Architect. 
+Synthesize the provided segment summaries into a cohesive, high-density "Master Strategic Briefing".
+
+STRICT CONSTRAINT: Total length must be between 1200 and 1800 words for a comprehensive deep-dive.
+
+Synthesis Strategy:
+1. ELIMINATE REDUNDANCY: Group similar findings from different segments together.
+2. LOGICAL MAPPING: Establish clear connections between methodology, results, and implications across the entire document.
+3. STRUCTURE: Use professional H2/H3 headers.
+4. TARGET SECTIONS: 
+   - Main Themes & Scope
+   - Technical Methodology & Contributions
+   - Primary Findings & Evidence
+   - Critical Implications & "So What?" analysis."""
 
 SUMMARY_REDUCE_PROMPT = ChatPromptTemplate.from_messages([
     ("system", SUMMARY_REDUCE_SYSTEM),
-    ("human", "Summaries:\n{text}\n\nGenerate the briefing.")
+    ("human", "Segment Summaries:\n{text}\n\nExecute the Master Strategic Briefing Summary:")
 ])
 
 PODCAST_AUDIO_SYSTEM = """You are producing a podcast script. 
@@ -123,9 +134,8 @@ PODCAST_AUDIO_PROMPT = ChatPromptTemplate.from_messages([
 ])
 
 
-
 # Knowledge Graph Prompt
-GRAPH_TEMPLATE = """You are an expert at visualizing complex academic information. 
+GRAPH_SYSTEM = """You are an expert at visualizing complex knowledge information. 
 Your goal is to extract a DEEP hierarchical structure and key relationships from the provided text and represent them as a CLEAN, multi-level Knowledge Graph using DOT syntax.
 
 CRITICAL INSTRUCTIONS:
@@ -135,7 +145,7 @@ CRITICAL INSTRUCTIONS:
 4. DESCRIPTIVE RELATIONSHIPS: Every edge MUST have a unique, descriptive label (e.g., "implements", "results in", "validates"). AVOID using the same generic label like "includes" for multiple edges in the same branch.
 5. AVOID SPIDER WEBS: Focus on hierarchical flow (Root -> Child -> Grandchild) rather than lateral cross-connections.
 6. HIERARCHICAL LAYOUT: Use 'rankdir=LR' (Left-to-Right).
-7. CONCISE LABELS: Keep node names and labels short (1-3 words).
+7. CONCISE LABELS: Keep node names and labels short (less than 5 words).
 
 Output ONLY the raw DOT code. No markdown code blocks.
 
@@ -154,9 +164,14 @@ digraph G {{
     "Category B" -> "Step 1" [label="baseline"];
     "Step 1" -> "Validation Method" [label="criteria"];
     "Validation Method" -> "Metric X" [label="output"];
-}}
+}}"""
 
-Text to Analyze:
+GRAPH_HUMAN = """Text to Analyze:
 {text}
 
 DOT Code:"""
+
+GRAPH_PROMPT = ChatPromptTemplate.from_messages([
+    ("system", GRAPH_SYSTEM),
+    ("human", GRAPH_HUMAN)
+])