import streamlit as st
from supabase import create_client
from google import genai
import os
# Page config
st.set_page_config(
page_title="Audio RAG",
page_icon="🎧",
layout="wide"
)
# Custom CSS
st.markdown("""
<style>
.file-card {
padding: 10px;
border-radius: 8px;
margin-bottom: 8px;
cursor: pointer;
}
.file-card:hover {
background-color: #f0f2f6;
}
.summary-container {
background-color: #f8f9fa;
padding: 20px;
border-radius: 10px;
border-left: 4px solid #4CAF50;
}
.search-result {
background-color: #ffffff;
padding: 15px;
border-radius: 8px;
border: 1px solid #e0e0e0;
margin-bottom: 10px;
}
.similarity-badge {
background-color: #e3f2fd;
padding: 2px 8px;
border-radius: 12px;
font-size: 12px;
}
</style>
""", unsafe_allow_html=True)
# Initialize clients
@st.cache_resource
def init_clients():
supabase = create_client(
os.environ.get("SUPABASE_URL", st.secrets.get("SUPABASE_URL", "")),
os.environ.get("SUPABASE_SERVICE_KEY", st.secrets.get("SUPABASE_SERVICE_KEY", ""))
)
ai = genai.Client(
api_key=os.environ.get("GEMINI_API_KEY", st.secrets.get("GEMINI_API_KEY", ""))
)
return supabase, ai
supabase, ai = init_clients()
# Helper functions
def generate_embedding(text: str) -> list:
response = ai.models.embed_content(
model="text-embedding-004",
contents=text,
)
return response.embeddings[0].values
def search_transcripts(query: str, match_count: int = 5) -> list:
query_vector = generate_embedding(query)
result = supabase.rpc("search_transcripts", {
"query_embedding": query_vector,
"match_threshold": 0.3,
"match_count": match_count,
}).execute()
return result.data or []
def generate_search_insights(query: str, results: list) -> str:
if not results:
return ""
# Combine relevant content
context = "\n\n---\n\n".join([
f"From '{r['source_file']}':\n{r['content']}"
for r in results[:5]
])
response = ai.models.generate_content(
model="gemini-2.5-flash",
contents=f"""Based on the user's question and the relevant audio transcript segments below, provide a helpful AI-generated answer.
**User's Question:** {query}
**Relevant Transcript Segments:**
{context}
**Instructions:**
- Directly answer the user's question using information from the transcripts
- Be concise but comprehensive
- If the transcripts don't fully answer the question, say what information is available
- Use bullet points for clarity when appropriate
- Mention which audio file(s) the information comes from
Provide your response in markdown format."""
)
return response.text
def get_all_files() -> list:
result = supabase.from_("transcripts").select("source_file, created_at").execute()
if not result.data:
return []
file_map = {}
for row in result.data:
if row["source_file"] not in file_map:
file_map[row["source_file"]] = {"count": 0, "date": row["created_at"]}
file_map[row["source_file"]]["count"] += 1
return [{"name": k, "chunks": v["count"], "date": v["date"]} for k, v in file_map.items()]
def get_full_transcript(source_file: str) -> str:
result = supabase.from_("transcripts").select("content").eq("source_file", source_file).order("created_at").execute()
if not result.data:
return ""
return " ".join([r["content"] for r in result.data])
def summarize_transcript(source_file: str) -> str:
transcript = get_full_transcript(source_file)
if not transcript:
return "No transcript found."
response = ai.models.generate_content(
model="gemini-2.5-flash",
contents=f"""Please provide a comprehensive summary of the following transcript in markdown format.
Structure it with:
## Overview
A brief 2-3 sentence overview
## Key Topics
- Bullet points of main topics
## Key Takeaways
- Important points and conclusions
## Action Items (if any)
- Any mentioned next steps or recommendations
Transcript:
{transcript}"""
)
return response.text
# Initialize session state
if "active_view" not in st.session_state:
st.session_state.active_view = "search"
if "selected_file" not in st.session_state:
st.session_state.selected_file = None
if "current_summary" not in st.session_state:
st.session_state.current_summary = None
# Get files
files = get_all_files()
# Sidebar - File list
with st.sidebar:
st.header("🎧 Audio RAG")
st.caption("Semantic search over audio transcriptions")
st.divider()
# Navigation
st.subheader("📁 Your Audio Files")
if files:
for f in files:
col1, col2 = st.columns([3, 1])
with col1:
display_name = f["name"][:25] + "..." if len(f["name"]) > 25 else f["name"]
if st.button(f"🎵 {display_name}", key=f"file_{f['name']}", use_container_width=True):
st.session_state.selected_file = f["name"]
st.session_state.active_view = "file_detail"
st.session_state.current_summary = None
with col2:
st.caption(f"{f['chunks']} chunks")
else:
st.info("No audio files yet")
st.divider()
# Quick actions
if st.button("🔍 Search", use_container_width=True):
st.session_state.active_view = "search"
st.session_state.selected_file = None
# Main content
if st.session_state.active_view == "search":
st.title("🔍 Search Your Audio")
st.markdown("Find specific topics, quotes, or discussions across all your transcriptions")
# Search input
col1, col2 = st.columns([4, 1])
with col1:
query = st.text_input("", placeholder="What was discussed about AI evaluation?", label_visibility="collapsed")
with col2:
num_results = st.selectbox("Results", [5, 10, 15, 20], label_visibility="collapsed")
if query:
with st.spinner("Searching your audio files..."):
results = search_transcripts(query, num_results)
if results:
# AI Insights section
st.markdown("### 🤖 AI Answer")
with st.spinner("Generating insights..."):
insights = generate_search_insights(query, results)
st.markdown(f"""
<div class="summary-container">
{insights}
</div>
""", unsafe_allow_html=True)
st.markdown(f"### 📚 Source Segments ({len(results)} found)")
st.caption("These are the transcript segments used to generate the answer above")
for i, r in enumerate(results):
with st.expander(f"📄 {r['source_file']} — {r['similarity']:.0%} match"):
st.markdown(f"> {r['content']}")
else:
st.warning("No results found. Try different keywords.")
else:
# Show recent files as suggestions
if files:
st.markdown("### 💡 Or explore your files")
cols = st.columns(min(3, len(files)))
for i, f in enumerate(files[:3]):
with cols[i]:
if st.button(f"📂 {f['name'][:20]}...", key=f"quick_{f['name']}", use_container_width=True):
st.session_state.selected_file = f["name"]
st.session_state.active_view = "file_detail"
st.rerun()
elif st.session_state.active_view == "file_detail" and st.session_state.selected_file:
selected = st.session_state.selected_file
# Header
st.title(f"🎵 {selected}")
# File info
file_info = next((f for f in files if f["name"] == selected), None)
if file_info:
col1, col2, col3 = st.columns(3)
with col1:
st.metric("Chunks", file_info["chunks"])
with col2:
st.metric("Added", file_info["date"][:10])
with col3:
chars = len(get_full_transcript(selected))
st.metric("Characters", f"{chars:,}")
st.divider()
# Action tabs
tab1, tab2 = st.tabs(["📝 Summary", "📄 Full Transcript"])
with tab1:
if st.button("✨ Generate AI Summary", type="primary", use_container_width=True):
with st.spinner("Generating summary with Gemini..."):
st.session_state.current_summary = summarize_transcript(selected)
if st.session_state.current_summary:
st.markdown('<div class="summary-container">', unsafe_allow_html=True)
st.markdown(st.session_state.current_summary)
st.markdown('</div>', unsafe_allow_html=True)
# Download button
st.download_button(
"📥 Download Summary",
st.session_state.current_summary,
file_name=f"{selected}_summary.md",
mime="text/markdown"
)
else:
st.info("Click the button above to generate an AI-powered summary of this audio file.")
with tab2:
transcript = get_full_transcript(selected)
st.text_area("", transcript, height=500, label_visibility="collapsed")
st.download_button(
"📥 Download Transcript",
transcript,
file_name=f"{selected}_transcript.txt",
mime="text/plain"
)
# Footer
st.sidebar.markdown("---")
st.sidebar.markdown("[GitHub](https://github.com/matheusslg/mcp-audio-rag) • Built with Streamlit")
