feat: prose completion api

src/config/agents.py

@@ -15,4 +15,5 @@ AGENT_LLM_MAP: dict[str, LLMType] = {
     "reporter": "basic",
     "podcast_script_writer": "basic",
     "ppt_composer": "basic",
+    "prose_writer": "basic",
 }

src/prose/graph/builder.py

@@ -0,0 +1,67 @@
+# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# SPDX-License-Identifier: MIT
+
+import asyncio
+import logging
+from langgraph.graph import END, START, StateGraph
+
+from src.prose.graph.prose_continue_node import prose_continue_node
+from src.prose.graph.prose_fix_node import prose_fix_node
+from src.prose.graph.prose_improve_node import prose_improve_node
+from src.prose.graph.prose_longer_node import prose_longer_node
+from src.prose.graph.prose_shorter_node import prose_shorter_node
+from src.prose.graph.prose_zap_node import prose_zap_node
+from src.prose.graph.state import ProseState
+
+
+def optional_node(state: ProseState):
+    return state["option"]
+
+
+def build_graph():
+    """Build and return the ppt workflow graph."""
+    # build state graph
+    builder = StateGraph(ProseState)
+    builder.add_node("prose_continue", prose_continue_node)
+    builder.add_node("prose_improve", prose_improve_node)
+    builder.add_node("prose_shorter", prose_shorter_node)
+    builder.add_node("prose_longer", prose_longer_node)
+    builder.add_node("prose_fix", prose_fix_node)
+    builder.add_node("prose_zap", prose_zap_node)
+    builder.add_conditional_edges(
+        START,
+        optional_node,
+        {
+            "continue": "prose_continue",
+            "improve": "prose_improve",
+            "shorter": "prose_shorter",
+            "longer": "prose_longer",
+            "fix": "prose_fix",
+            "zap": "prose_zap",
+        },
+        END,
+    )
+    return builder.compile()
+
+
+async def _test_workflow():
+    workflow = build_graph()
+    events = workflow.astream(
+        {
+            "content": "The weather in Beijing is sunny",
+            "option": "continue",
+        },
+        stream_mode="messages",
+        subgraphs=True,
+    )
+    async for node, event in events:
+        e = event[0]
+        print({"id": e.id, "object": "chat.completion.chunk", "content": e.content})
+
+
+if __name__ == "__main__":
+    from dotenv import load_dotenv
+
+    load_dotenv()
+    logging.basicConfig(level=logging.INFO)
+    asyncio.run(_test_workflow())

src/prose/graph/prose_continue_node.py

@@ -0,0 +1,31 @@
+# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# SPDX-License-Identifier: MIT
+
+import logging
+
+from langchain.schema import HumanMessage, SystemMessage
+
+from src.config.agents import AGENT_LLM_MAP
+from src.llms.llm import get_llm_by_type
+from src.prose.graph.state import ProseState
+
+logger = logging.getLogger(__name__)
+
+
+def prose_continue_node(state: ProseState):
+    logger.info("Generating prose continue content...")
+    model = get_llm_by_type(AGENT_LLM_MAP["prose_writer"])
+    prose_content = model.invoke(
+        [
+            SystemMessage(
+                content="""
+You are an AI writing assistant that continues existing text based on context from prior text.
+- Give more weight/priority to the later characters than the beginning ones.
+- Limit your response to no more than 200 characters, but make sure to construct complete sentences.
+- Use Markdown formatting when appropriate
+"""
+            ),
+            HumanMessage(content=state["content"]),
+        ],
+    )
+    return {"output": prose_content.content}

src/prose/graph/prose_fix_node.py

@@ -0,0 +1,32 @@
+# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# SPDX-License-Identifier: MIT
+
+import logging
+
+from langchain.schema import HumanMessage, SystemMessage
+
+from src.config.agents import AGENT_LLM_MAP
+from src.llms.llm import get_llm_by_type
+from src.prose.graph.state import ProseState
+
+logger = logging.getLogger(__name__)
+
+
+def prose_fix_node(state: ProseState):
+    logger.info("Generating prose fix content...")
+    model = get_llm_by_type(AGENT_LLM_MAP["prose_writer"])
+    prose_content = model.invoke(
+        [
+            SystemMessage(
+                content="""
+You are an AI writing assistant that fixes grammar and spelling errors in existing text. 
+- Limit your response to no more than 200 characters, but make sure to construct complete sentences.
+- Use Markdown formatting when appropriate.
+- If the text is already correct, just return the original text.
+"""
+            ),
+            HumanMessage(content=f"The existing text is: {state['content']}"),
+        ],
+    )
+    logger.info(f"prose_content: {prose_content}")
+    return {"output": prose_content.content}

src/prose/graph/prose_improve_node.py

@@ -0,0 +1,31 @@
+# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# SPDX-License-Identifier: MIT
+
+import logging
+
+from langchain.schema import HumanMessage, SystemMessage
+
+from src.config.agents import AGENT_LLM_MAP
+from src.llms.llm import get_llm_by_type
+from src.prose.graph.state import ProseState
+
+logger = logging.getLogger(__name__)
+
+
+def prose_improve_node(state: ProseState):
+    logger.info("Generating prose improve content...")
+    model = get_llm_by_type(AGENT_LLM_MAP["prose_writer"])
+    prose_content = model.invoke(
+        [
+            SystemMessage(
+                content="""
+You are an AI writing assistant that improves existing text.
+- Limit your response to no more than 200 characters, but make sure to construct complete sentences.
+- Use Markdown formatting when appropriate.
+"""
+            ),
+            HumanMessage(content=f"The existing text is: {state['content']}"),
+        ],
+    )
+    logger.info(f"prose_content: {prose_content}")
+    return {"output": prose_content.content}

src/prose/graph/prose_longer_node.py

@@ -0,0 +1,30 @@
+# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# SPDX-License-Identifier: MIT
+
+import logging
+
+from langchain.schema import HumanMessage, SystemMessage
+
+from src.config.agents import AGENT_LLM_MAP
+from src.llms.llm import get_llm_by_type
+from src.prose.graph.state import ProseState
+
+logger = logging.getLogger(__name__)
+
+
+def prose_longer_node(state: ProseState):
+    logger.info("Generating prose longer content...")
+    model = get_llm_by_type(AGENT_LLM_MAP["prose_writer"])
+    prose_content = model.invoke(
+        [
+            SystemMessage(
+                content="""
+You are an AI writing assistant that lengthens existing text.
+- Use Markdown formatting when appropriate.
+"""
+            ),
+            HumanMessage(content=f"The existing text is: {state['content']}"),
+        ],
+    )
+    logger.info(f"prose_content: {prose_content}")
+    return {"output": prose_content.content}

src/prose/graph/prose_shorter_node.py

@@ -0,0 +1,30 @@
+# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# SPDX-License-Identifier: MIT
+
+import logging
+
+from langchain.schema import HumanMessage, SystemMessage
+
+from src.config.agents import AGENT_LLM_MAP
+from src.llms.llm import get_llm_by_type
+from src.prose.graph.state import ProseState
+
+logger = logging.getLogger(__name__)
+
+
+def prose_shorter_node(state: ProseState):
+    logger.info("Generating prose shorter content...")
+    model = get_llm_by_type(AGENT_LLM_MAP["prose_writer"])
+    prose_content = model.invoke(
+        [
+            SystemMessage(
+                content="""
+You are an AI writing assistant that shortens existing text.
+- Use Markdown formatting when appropriate.
+"""
+            ),
+            HumanMessage(content=f"The existing text is: {state['content']}"),
+        ],
+    )
+    logger.info(f"prose_content: {prose_content}")
+    return {"output": prose_content.content}

src/prose/graph/prose_zap_node.py

@@ -0,0 +1,33 @@
+# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# SPDX-License-Identifier: MIT
+
+import logging
+
+from langchain.schema import HumanMessage, SystemMessage
+
+from src.config.agents import AGENT_LLM_MAP
+from src.llms.llm import get_llm_by_type
+from src.prose.graph.state import ProseState
+
+logger = logging.getLogger(__name__)
+
+
+def prose_zap_node(state: ProseState):
+    logger.info("Generating prose zap content...")
+    model = get_llm_by_type(AGENT_LLM_MAP["prose_writer"])
+    prose_content = model.invoke(
+        [
+            SystemMessage(
+                content="""
+You area an AI writing assistant that generates text based on a prompt. 
+- You take an input from the user and a command for manipulating the text."
+- Use Markdown formatting when appropriate.
+"""
+            ),
+            HumanMessage(
+                content=f"For this text: {state['content']}.\nYou have to respect the command: {state['command']}"
+            ),
+        ],
+    )
+    logger.info(f"prose_content: {prose_content}")
+    return {"output": prose_content.content}

src/prose/graph/state.py

@@ -0,0 +1,20 @@
+# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# SPDX-License-Identifier: MIT
+
+from langgraph.graph import MessagesState
+
+
+class ProseState(MessagesState):
+    """State for the prose generation."""
+
+    # The content of the prose
+    content: str = ""
+
+    # Prose writer option: continue, improve, shorter, longer, fix, zap
+    option: str = ""
+
+    # The user custom command for the prose writer
+    command: str = ""
+
+    # Output
+    output: str = ""

src/server/app.py

@@ -17,11 +17,13 @@ from langgraph.types import Command
 from src.graph.builder import build_graph_with_memory
 from src.podcast.graph.builder import build_graph as build_podcast_graph
 from src.ppt.graph.builder import build_graph as build_ppt_graph
+from src.prose.graph.builder import build_graph as build_prose_graph
 from src.server.chat_request import (
     ChatMessage,
     ChatRequest,
     GeneratePodcastRequest,
     GeneratePPTRequest,
+    GenerateProseRequest,
     TTSRequest,
 )
 from src.server.mcp_request import MCPServerMetadataRequest, MCPServerMetadataResponse
@@ -254,6 +256,29 @@ async def generate_ppt(request: GeneratePPTRequest):
         raise HTTPException(status_code=500, detail=str(e))
 
 
+@app.post("/api/prose/generate")
+async def generate_prose(request: GenerateProseRequest):
+    try:
+        logger.info(f"Generating prose for prompt: {request.prompt}")
+        workflow = build_prose_graph()
+        events = workflow.astream(
+            {
+                "content": request.prompt,
+                "option": request.option,
+                "command": request.command,
+            },
+            stream_mode="messages",
+            subgraphs=True,
+        )
+        return StreamingResponse(
+            (f"data: {event[0].content}\n\n" async for _, event in events),
+            media_type="text/event-stream",
+        )
+    except Exception as e:
+        logger.exception(f"Error occurred during prose generation: {str(e)}")
+        raise HTTPException(status_code=500, detail=str(e))
+
+
 @app.post("/api/mcp/server/metadata", response_model=MCPServerMetadataResponse)
 async def mcp_server_metadata(request: MCPServerMetadataRequest):
     """Get information about an MCP server."""

src/server/chat_request.py

@@ -74,3 +74,11 @@ class GeneratePodcastRequest(BaseModel):
 
 class GeneratePPTRequest(BaseModel):
     content: str = Field(..., description="The content of the ppt")
+
+
+class GenerateProseRequest(BaseModel):
+    prompt: str = Field(..., description="The content of the prose")
+    option: str = Field(..., description="The option of the prose writer")
+    command: Optional[str] = Field(
+        "", description="The user custom command of the prose writer"
+    )