如何用 Streamlit 和 Snowflake Cortex 搭建语音助手应用 ｜技术实践

git clone https://github.com/streamlit/30DaysOfAI.gitcd 30DaysOfAI/app

CREATE DATABASE IF NOT EXISTS RAG_DB;CREATE SCHEMA IF NOT EXISTS RAG_DB.RAG_SCHEMA;DROP STAGE IF EXISTS RAG_DB.RAG_SCHEMA.VOICE_AUDIO;CREATE STAGE RAG_DB.RAG_SCHEMA.VOICE_AUDIO    DIRECTORY = ( ENABLE = true )    ENCRYPTION = ( TYPE = 'SNOWFLAKE_SSE' );

import streamlit as stimport jsonfrom snowflake.snowpark.functions import ai_completeimport ioimport timeimport hashlibtry:    from snowflake.snowpark.context import get_active_session    session = get_active_session()except:    from snowflake.snowpark import Session    session = Session.builder.configs(st.secrets["connections"]["snowflake"]).create()def call_llm(prompt_text: str) -> str:    df = session.range(1).select(        ai_complete(model="claude-3-5-sonnet", prompt=prompt_text).alias("response")    )    response_raw = df.collect()[0][0]    response_json = json.loads(response_raw)    if isinstance(response_json, dict):        return response_json.get("choices", [{}])[0].get("messages", "")    return str(response_json)if "voice_messages" not in st.session_state:    st.session_state.voice_messages = []if len(st.session_state.voice_messages) == 0:    st.session_state.voice_messages = [        {"role": "assistant", "content": "Hello! :material/waving_hand: I'm your voice-enabled AI assistant. Click the microphone button to record a message, and I'll respond to you!"}    ]if "voice_database" not in st.session_state:    st.session_state.voice_database = "RAG_DB"    st.session_state.voice_schema = "RAG_SCHEMA"if "processed_audio_id" not in st.session_state:    st.session_state.processed_audio_id = None

database = st.session_state.voice_databaseschema = st.session_state.voice_schemafull_stage_name = f"{database}.{schema}.VOICE_AUDIO"stage_name = f"@{full_stage_name}"with st.sidebar:    st.title(":material/record_voice_over: Voice-Enabled Assistant")    st.write("Talk to your AI assistant using voice input!")        st.header(":material/settings: Settings")        with st.expander("Stage Status", expanded=False):        try:            stage_info = session.sql(f"SHOW STAGES LIKE 'VOICE_AUDIO' IN SCHEMA {database}.{schema}").collect()            if stage_info:                session.sql(f"DROP STAGE IF EXISTS {full_stage_name}").collect()            session.sql(f"""            CREATE STAGE {full_stage_name}                DIRECTORY = ( ENABLE = true )                ENCRYPTION = ( TYPE = 'SNOWFLAKE_SSE' )            """).collect()            st.success(":material/check_box: Audio stage ready (server-side encrypted)")        except Exception as e:            st.error(f":material/cancel: Could not create stage")        if st.button(":material/delete: Clear Chat"):        st.session_state.voice_messages = [            {"role": "assistant", "content": "Hello! :material/waving_hand: I'm your voice-enabled AI assistant. Click the microphone button to record a message, and I'll respond to you!"}        ]        st.rerun()

st.subheader(":material/voice_chat: Conversation")audio = st.audio_input(":material/mic: Click to record")for msg in st.session_state.voice_messages:    with st.chat_message(msg["role"]):        st.markdown(msg["content"])status_container = st.container()if audio is not None:    audio_bytes = audio.read()    audio_hash = hashlib.md5(audio_bytes).hexdigest()        if audio_hash != st.session_state.processed_audio_id:        st.session_state.processed_audio_id = audio_hash                with status_container:            transcript = None            with st.spinner(":material/mic: Transcribing audio..."):                try:                    timestamp = int(time.time())                    filename = f"audio_{timestamp}.wav"                                        audio_stream = io.BytesIO(audio_bytes)                    full_stage_path = f"{stage_name}/{filename}"                                        session.file.put_stream(                        audio_stream,                        full_stage_path,                        overwrite=True,                        auto_compress=False                    )                                        safe_file_name = filename.replace("'", "''")                                        sql_query = f"""                    SELECT SNOWFLAKE.CORTEX.AI_TRANSCRIBE(                        TO_FILE('{stage_name}', '{safe_file_name}')                    ) as transcript                    """                                        result_rows = session.sql(sql_query).collect()                                        if result_rows:                        json_string = result_rows[0]['TRANSCRIPT']                        transcript_data = json.loads(json_string)                        transcript = transcript_data.get("text", "")                                                if transcript:                            st.session_state.voice_messages.append({                                "role": "user",                                "content": transcript                            })                                except Exception as e:                    st.error(f"Error during transcription: {str(e)}")

            if transcript:                with st.spinner(":material/smart_toy: Generating response..."):                    conversation_context = "You are a friendly voice assistant. Keep responses short and conversational.\n\nConversation history:\n"                                        history_messages = [msg for msg in st.session_state.voice_messages[:-1]                                        if not (msg["role"] == "assistant" and "Click the microphone" in msg["content"])]                                        for msg in history_messages:                        role = "User" if msg["role"] == "user" else "Assistant"                        conversation_context += f"{role}: {msg['content']}\n"                                        conversation_context += f"\nUser: {transcript}\n\nAssistant:"                                        response = call_llm(conversation_context)                                        st.session_state.voice_messages.append({                        "role": "assistant",                        "content": response                    })                                try:                    session.sql(f"REMOVE {stage_name}/{safe_file_name}").collect()                except:                    pass                                st.rerun()else:    st.session_state.processed_audio_id = None

import streamlit as stimport jsonfrom snowflake.snowpark.functions import ai_completeimport ioimport timeimport hashlibtry:    from snowflake.snowpark.context import get_active_session    session = get_active_session()except:    from snowflake.snowpark import Session    session = Session.builder.configs(st.secrets["connections"]["snowflake"]).create()def call_llm(prompt_text: str) -> str:    """Call Snowflake Cortex LLM."""    df = session.range(1).select(        ai_complete(model="claude-3-5-sonnet", prompt=prompt_text).alias("response")    )    response_raw = df.collect()[0][0]    response_json = json.loads(response_raw)    if isinstance(response_json, dict):        return response_json.get("choices", [{}])[0].get("messages", "")    return str(response_json)if "voice_messages" not in st.session_state:    st.session_state.voice_messages = []if len(st.session_state.voice_messages) == 0:    st.session_state.voice_messages = [        {            "role": "assistant",            "content": "Hello! :material/waving_hand: I'm your voice-enabled AI assistant. Click the microphone button to record a message, and I'll respond to you!"        }    ]if "voice_database" not in st.session_state:    st.session_state.voice_database = "RAG_DB"    st.session_state.voice_schema = "RAG_SCHEMA"if "processed_audio_id" not in st.session_state:    st.session_state.processed_audio_id = Nonedatabase = st.session_state.voice_databaseschema = st.session_state.voice_schemafull_stage_name = f"{database}.{schema}.VOICE_AUDIO"stage_name = f"@{full_stage_name}"with st.sidebar:    st.title(":material/record_voice_over: Voice-Enabled Assistant")    st.write("Talk to your AI assistant using voice input!")        st.header(":material/settings: Settings")        with st.expander("Database Configuration", expanded=False):        database = st.text_input("Database", value=st.session_state.voice_database, key="db_input")        schema = st.text_input("Schema", value=st.session_state.voice_schema, key="schema_input")                st.session_state.voice_database = database        st.session_state.voice_schema = schema                st.caption(f"Stage: `{database}.{schema}.VOICE_AUDIO`")        st.caption(":material/edit_note: Stage uses server-side encryption (required for AI_TRANSCRIBE)")                if st.button(":material/autorenew: Recreate Stage", help="Drop and recreate the stage with correct encryption"):            try:                full_stage = f"{database}.{schema}.VOICE_AUDIO"                session.sql(f"DROP STAGE IF EXISTS {full_stage}").collect()                session.sql(f"""                    CREATE STAGE {full_stage}                        DIRECTORY = ( ENABLE = true )                        ENCRYPTION = ( TYPE = 'SNOWFLAKE_SSE' )                """).collect()                st.success(f":material/check_circle: Stage recreated successfully!")                st.rerun()            except Exception as e:                st.error(f"Failed to recreate stage: {str(e)}")        with st.expander("Stage Status", expanded=False):        database = st.session_state.voice_database        schema = st.session_state.voice_schema        full_stage_name = f"{database}.{schema}.VOICE_AUDIO"                try:            stage_info = session.sql(f"SHOW STAGES LIKE 'VOICE_AUDIO' IN SCHEMA {database}.{schema}").collect()                        if stage_info:                st.info(f":material/autorenew: Recreating stage with server-side encryption...")                session.sql(f"DROP STAGE IF EXISTS {full_stage_name}").collect()                        session.sql(f"""                CREATE STAGE {full_stage_name}                    DIRECTORY = ( ENABLE = true )                    ENCRYPTION = ( TYPE = 'SNOWFLAKE_SSE' )            """).collect()            st.success(f":material/check_box: Audio stage ready (server-side encrypted)")                    except Exception as e:            st.error(f":material/cancel: Could not create stage")        if st.button(":material/delete: Clear Chat"):        st.session_state.voice_messages = [            {                "role": "assistant",                "content": "Hello! :material/waving_hand: I'm your voice-enabled AI assistant. Click the microphone button to record a message, and I'll respond to you!"            }        ]        st.rerun()st.subheader(":material/voice_chat: Conversation")audio = st.audio_input(":material/mic: Click to record")for msg in st.session_state.voice_messages:    with st.chat_message(msg["role"]):        st.markdown(msg["content"])status_container = st.container()if audio is not None:    audio_bytes = audio.read()    audio_hash = hashlib.md5(audio_bytes).hexdigest()        if audio_hash != st.session_state.processed_audio_id:        st.session_state.processed_audio_id = audio_hash                with status_container:            transcript = None            with st.spinner(":material/mic: Transcribing audio..."):                try:                    timestamp = int(time.time())                    filename = f"audio_{timestamp}.wav"                                        audio_stream = io.BytesIO(audio_bytes)                    full_stage_path = f"{stage_name}/{filename}"                                        session.file.put_stream(                        audio_stream,                        full_stage_path,                        overwrite=True,                        auto_compress=False                    )                                        safe_file_name = filename.replace("'", "''")                                        sql_query = f"""                    SELECT SNOWFLAKE.CORTEX.AI_TRANSCRIBE(                        TO_FILE('{stage_name}', '{safe_file_name}')                    ) as transcript                    """                                        result_rows = session.sql(sql_query).collect()                                        if result_rows:                        json_string = result_rows[0]['TRANSCRIPT']                        transcript_data = json.loads(json_string)                        transcript = transcript_data.get("text", "")                                                if transcript:                            st.session_state.voice_messages.append({                                "role": "user",                                "content": transcript                            })                        else:                            st.error("Transcription returned no text.")                            st.json(transcript_data)                    else:                        st.error("Transcription query returned no results.")                                except Exception as e:                    st.error(f"Error during transcription: {str(e)}")                        if transcript:                with st.spinner(":material/smart_toy: Generating response..."):                    conversation_context = "You are a friendly voice assistant. Keep responses short and conversational.\n\nConversation history:\n"                                        history_messages = st.session_state.voice_messages[:-1] if len(st.session_state.voice_messages) > 1 else []                                        history_messages = [msg for msg in history_messages if not (msg["role"] == "assistant" and "Click the microphone button" in msg["content"])]                                        for msg in history_messages:                        role = "User" if msg["role"] == "user" else "Assistant"                        conversation_context += f"{role}: {msg['content']}\n"                                        conversation_context += f"\nUser: {transcript}\n\nAssistant:"                                        response = call_llm(conversation_context)                                        st.session_state.voice_messages.append({                        "role": "assistant",                        "content": response                    })                                try:                    session.sql(f"REMOVE {stage_name}/{safe_file_name}").collect()                except:                    pass                                st.rerun()else:    st.session_state.processed_audio_id = Nonest.divider()st.caption("Day 25: Voice Interface | 30 Days of AI")