Spaces:
Sleeping
Sleeping
mriusero commited on
Commit ·
42500ca
1
Parent(s): c4d08fc
core: refacto
Browse files- src/agent/{inference.py → mistral_agent.py} +0 -0
- src/agent/stream.py +125 -0
- src/ui/sidebar.py +1 -125
src/agent/{inference.py → mistral_agent.py}
RENAMED
|
File without changes
|
src/agent/stream.py
ADDED
|
@@ -0,0 +1,125 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from gradio import ChatMessage
|
| 2 |
+
|
| 3 |
+
from src.agent.mistral_agent import MistralAgent
|
| 4 |
+
|
| 5 |
+
agent = MistralAgent()
|
| 6 |
+
|
| 7 |
+
with open("./prompt.md", encoding="utf-8") as f:
|
| 8 |
+
SYSTEM_PROMPT = f.read()
|
| 9 |
+
|
| 10 |
+
async def respond(message, history=None):
|
| 11 |
+
"""
|
| 12 |
+
Respond to a user message using the Mistral agent.
|
| 13 |
+
"""
|
| 14 |
+
if history is None:
|
| 15 |
+
history = []
|
| 16 |
+
|
| 17 |
+
history.append(ChatMessage(role="user", content=message))
|
| 18 |
+
history.append(ChatMessage(role="assistant", content="", metadata={"title": "Thinking", "status": "pending"}))
|
| 19 |
+
yield history
|
| 20 |
+
|
| 21 |
+
messages = [
|
| 22 |
+
{"role": "system", "content": SYSTEM_PROMPT},
|
| 23 |
+
{"role": "user", "content": message},
|
| 24 |
+
{
|
| 25 |
+
"role": "assistant",
|
| 26 |
+
"content": "THINKING: Let's tackle this problem, ",
|
| 27 |
+
"prefix": True,
|
| 28 |
+
},
|
| 29 |
+
]
|
| 30 |
+
payload = {
|
| 31 |
+
"agent_id": agent.agent_id,
|
| 32 |
+
"messages": messages,
|
| 33 |
+
"stream": True,
|
| 34 |
+
"max_tokens": None,
|
| 35 |
+
"tools": agent.tools,
|
| 36 |
+
"tool_choice": "auto",
|
| 37 |
+
"presence_penalty": 0,
|
| 38 |
+
"frequency_penalty": 0,
|
| 39 |
+
"n": 1
|
| 40 |
+
}
|
| 41 |
+
response = await agent.client.agents.stream_async(**payload)
|
| 42 |
+
|
| 43 |
+
full = ""
|
| 44 |
+
thinking = ""
|
| 45 |
+
tooling = ""
|
| 46 |
+
final = ""
|
| 47 |
+
|
| 48 |
+
current_phase = None # None | "thinking" | "tooling" | "final"
|
| 49 |
+
|
| 50 |
+
history[-1] = ChatMessage(role="assistant", content="", metadata={"title": "Thinking", "status": "pending"})
|
| 51 |
+
|
| 52 |
+
async for chunk in response:
|
| 53 |
+
delta = chunk.data.choices[0].delta
|
| 54 |
+
content = delta.content or ""
|
| 55 |
+
full += content
|
| 56 |
+
|
| 57 |
+
# Phase finale
|
| 58 |
+
if "FINAL ANSWER:" in full:
|
| 59 |
+
|
| 60 |
+
parts = full.split("FINAL ANSWER:", 1)
|
| 61 |
+
before_final = parts[0]
|
| 62 |
+
final = parts[1].strip()
|
| 63 |
+
|
| 64 |
+
if "TOOLING:" in before_final:
|
| 65 |
+
tooling = before_final.split("TOOLING:", 1)[1].strip()
|
| 66 |
+
else:
|
| 67 |
+
tooling = ""
|
| 68 |
+
|
| 69 |
+
if current_phase != "final":
|
| 70 |
+
if current_phase == "tooling":
|
| 71 |
+
history[-1] = ChatMessage(role="assistant", content=tooling, metadata={"title": "Tooling", "status": "done"})
|
| 72 |
+
elif current_phase == "thinking":
|
| 73 |
+
history[-1] = ChatMessage(role="assistant", content=thinking, metadata={"title": "Thinking", "status": "done"})
|
| 74 |
+
|
| 75 |
+
history.append(ChatMessage(role="assistant", content=final))
|
| 76 |
+
current_phase = "final"
|
| 77 |
+
yield history
|
| 78 |
+
|
| 79 |
+
# Phase outil
|
| 80 |
+
elif "TOOLING:" in full:
|
| 81 |
+
|
| 82 |
+
parts = full.split("TOOLING:", 1)
|
| 83 |
+
before_tooling = parts[0]
|
| 84 |
+
tooling = ""
|
| 85 |
+
|
| 86 |
+
if "THINKING:" in before_tooling:
|
| 87 |
+
thinking = before_tooling.split("THINKING:", 1)[1].strip()
|
| 88 |
+
else:
|
| 89 |
+
thinking = before_tooling.strip()
|
| 90 |
+
|
| 91 |
+
tooling = parts[1].strip()
|
| 92 |
+
|
| 93 |
+
if current_phase != "tooling":
|
| 94 |
+
if current_phase == "thinking":
|
| 95 |
+
history[-1] = ChatMessage(role="assistant", content=thinking,
|
| 96 |
+
metadata={"title": "Thinking", "status": "done"})
|
| 97 |
+
history.append(
|
| 98 |
+
ChatMessage(role="assistant", content=tooling, metadata={"title": "Tooling", "status": "pending"}))
|
| 99 |
+
current_phase = "tooling"
|
| 100 |
+
else:
|
| 101 |
+
history[-1] = ChatMessage(role="assistant", content=tooling,
|
| 102 |
+
metadata={"title": "Tooling", "status": "pending"})
|
| 103 |
+
yield history
|
| 104 |
+
|
| 105 |
+
# Phase réflexion
|
| 106 |
+
elif "THINKING:" in full or current_phase is None:
|
| 107 |
+
|
| 108 |
+
if "THINKING:" in full:
|
| 109 |
+
thinking = full.split("THINKING:", 1)[1].strip()
|
| 110 |
+
else:
|
| 111 |
+
thinking = full.strip()
|
| 112 |
+
|
| 113 |
+
if current_phase != "thinking":
|
| 114 |
+
history[-1] = ChatMessage(role="assistant", content=thinking, metadata={"title": "Thinking", "status": "pending"})
|
| 115 |
+
current_phase = "thinking"
|
| 116 |
+
else:
|
| 117 |
+
history[-1] = ChatMessage(role="assistant", content=thinking, metadata={"title": "Thinking", "status": "pending"})
|
| 118 |
+
yield history
|
| 119 |
+
|
| 120 |
+
if current_phase == "thinking":
|
| 121 |
+
history[-1] = ChatMessage(role="assistant", content=thinking, metadata={"title": "Thinking", "status": "done"})
|
| 122 |
+
elif current_phase == "tooling":
|
| 123 |
+
history[-1] = ChatMessage(role="assistant", content=tooling, metadata={"title": "Tooling", "status": "done"})
|
| 124 |
+
|
| 125 |
+
yield history
|
src/ui/sidebar.py
CHANGED
|
@@ -1,130 +1,6 @@
|
|
| 1 |
import gradio as gr
|
| 2 |
-
import json
|
| 3 |
-
from gradio import ChatMessage
|
| 4 |
|
| 5 |
-
from src.agent.
|
| 6 |
-
|
| 7 |
-
agent = MistralAgent()
|
| 8 |
-
|
| 9 |
-
with open("./prompt.md", encoding="utf-8") as f:
|
| 10 |
-
SYSTEM_PROMPT = f.read()
|
| 11 |
-
|
| 12 |
-
async def respond(message, history=None):
|
| 13 |
-
"""
|
| 14 |
-
Respond to a user message using the Mistral agent.
|
| 15 |
-
"""
|
| 16 |
-
if history is None:
|
| 17 |
-
history = []
|
| 18 |
-
|
| 19 |
-
history.append(ChatMessage(role="user", content=message))
|
| 20 |
-
history.append(ChatMessage(role="assistant", content="", metadata={"title": "Thinking", "status": "pending"}))
|
| 21 |
-
yield history
|
| 22 |
-
|
| 23 |
-
messages = [
|
| 24 |
-
{"role": "system", "content": SYSTEM_PROMPT},
|
| 25 |
-
{"role": "user", "content": message},
|
| 26 |
-
{
|
| 27 |
-
"role": "assistant",
|
| 28 |
-
"content": "THINKING: Let's tackle this problem, ",
|
| 29 |
-
"prefix": True,
|
| 30 |
-
},
|
| 31 |
-
]
|
| 32 |
-
payload = {
|
| 33 |
-
"agent_id": agent.agent_id,
|
| 34 |
-
"messages": messages,
|
| 35 |
-
"stream": True,
|
| 36 |
-
"max_tokens": None,
|
| 37 |
-
"tools": agent.tools,
|
| 38 |
-
"tool_choice": "auto",
|
| 39 |
-
"presence_penalty": 0,
|
| 40 |
-
"frequency_penalty": 0,
|
| 41 |
-
"n": 1
|
| 42 |
-
}
|
| 43 |
-
response = await agent.client.agents.stream_async(**payload)
|
| 44 |
-
|
| 45 |
-
full = ""
|
| 46 |
-
thinking = ""
|
| 47 |
-
tooling = ""
|
| 48 |
-
final = ""
|
| 49 |
-
|
| 50 |
-
current_phase = None # None | "thinking" | "tooling" | "final"
|
| 51 |
-
|
| 52 |
-
history[-1] = ChatMessage(role="assistant", content="", metadata={"title": "Thinking", "status": "pending"})
|
| 53 |
-
|
| 54 |
-
async for chunk in response:
|
| 55 |
-
delta = chunk.data.choices[0].delta
|
| 56 |
-
content = delta.content or ""
|
| 57 |
-
full += content
|
| 58 |
-
|
| 59 |
-
# Phase finale
|
| 60 |
-
if "FINAL ANSWER:" in full:
|
| 61 |
-
|
| 62 |
-
parts = full.split("FINAL ANSWER:", 1)
|
| 63 |
-
before_final = parts[0]
|
| 64 |
-
final = parts[1].strip()
|
| 65 |
-
|
| 66 |
-
if "TOOLING:" in before_final:
|
| 67 |
-
tooling = before_final.split("TOOLING:", 1)[1].strip()
|
| 68 |
-
else:
|
| 69 |
-
tooling = ""
|
| 70 |
-
|
| 71 |
-
if current_phase != "final":
|
| 72 |
-
if current_phase == "tooling":
|
| 73 |
-
history[-1] = ChatMessage(role="assistant", content=tooling, metadata={"title": "Tooling", "status": "done"})
|
| 74 |
-
elif current_phase == "thinking":
|
| 75 |
-
history[-1] = ChatMessage(role="assistant", content=thinking, metadata={"title": "Thinking", "status": "done"})
|
| 76 |
-
|
| 77 |
-
history.append(ChatMessage(role="assistant", content=final))
|
| 78 |
-
current_phase = "final"
|
| 79 |
-
yield history
|
| 80 |
-
|
| 81 |
-
# Phase outil
|
| 82 |
-
elif "TOOLING:" in full:
|
| 83 |
-
|
| 84 |
-
parts = full.split("TOOLING:", 1)
|
| 85 |
-
before_tooling = parts[0]
|
| 86 |
-
tooling = ""
|
| 87 |
-
|
| 88 |
-
if "THINKING:" in before_tooling:
|
| 89 |
-
thinking = before_tooling.split("THINKING:", 1)[1].strip()
|
| 90 |
-
else:
|
| 91 |
-
thinking = before_tooling.strip()
|
| 92 |
-
|
| 93 |
-
tooling = parts[1].strip()
|
| 94 |
-
|
| 95 |
-
if current_phase != "tooling":
|
| 96 |
-
if current_phase == "thinking":
|
| 97 |
-
history[-1] = ChatMessage(role="assistant", content=thinking,
|
| 98 |
-
metadata={"title": "Thinking", "status": "done"})
|
| 99 |
-
history.append(
|
| 100 |
-
ChatMessage(role="assistant", content=tooling, metadata={"title": "Tooling", "status": "pending"}))
|
| 101 |
-
current_phase = "tooling"
|
| 102 |
-
else:
|
| 103 |
-
history[-1] = ChatMessage(role="assistant", content=tooling,
|
| 104 |
-
metadata={"title": "Tooling", "status": "pending"})
|
| 105 |
-
yield history
|
| 106 |
-
|
| 107 |
-
# Phase réflexion
|
| 108 |
-
elif "THINKING:" in full or current_phase is None:
|
| 109 |
-
|
| 110 |
-
if "THINKING:" in full:
|
| 111 |
-
thinking = full.split("THINKING:", 1)[1].strip()
|
| 112 |
-
else:
|
| 113 |
-
thinking = full.strip()
|
| 114 |
-
|
| 115 |
-
if current_phase != "thinking":
|
| 116 |
-
history[-1] = ChatMessage(role="assistant", content=thinking, metadata={"title": "Thinking", "status": "pending"})
|
| 117 |
-
current_phase = "thinking"
|
| 118 |
-
else:
|
| 119 |
-
history[-1] = ChatMessage(role="assistant", content=thinking, metadata={"title": "Thinking", "status": "pending"})
|
| 120 |
-
yield history
|
| 121 |
-
|
| 122 |
-
if current_phase == "thinking":
|
| 123 |
-
history[-1] = ChatMessage(role="assistant", content=thinking, metadata={"title": "Thinking", "status": "done"})
|
| 124 |
-
elif current_phase == "tooling":
|
| 125 |
-
history[-1] = ChatMessage(role="assistant", content=tooling, metadata={"title": "Tooling", "status": "done"})
|
| 126 |
-
|
| 127 |
-
yield history
|
| 128 |
|
| 129 |
|
| 130 |
def sidebar_ui(state, width=700, visible=True):
|
|
|
|
| 1 |
import gradio as gr
|
|
|
|
|
|
|
| 2 |
|
| 3 |
+
from src.agent.stream import respond
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 4 |
|
| 5 |
|
| 6 |
def sidebar_ui(state, width=700, visible=True):
|