File size: 9,578 Bytes
458c8e2 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 |
# agent/agent_graph.py
from __future__ import annotations
import json
from typing import Any, Dict, List, Optional, TypedDict
from langchain_core.messages import AIMessage, HumanMessage, SystemMessage, ToolMessage
from langgraph.graph import END, START, StateGraph
from langgraph.prebuilt import ToolNode
from .runtime_ctx import set_df_payload, set_df_summary, set_sota_bundled
from .tools import (
tool_describe_step,
tool_inspect_dataset,
tool_list_steps,
tool_list_versions,
tool_propose_plan,
tool_reset_to_version,
tool_run_step,
tool_sota_preprocessing,
)
def _to_text(content: Any, limit: int = 4000) -> str:
"""Coerce any message content to a string that Gemini will accept."""
if content is None:
return ""
if isinstance(content, str):
return content
try:
s = json.dumps(content, default=str, ensure_ascii=False)
except Exception:
s = str(content)
# lightly truncate huge tool dumps
return (s[:limit] + " …") if len(s) > limit else s
def _sanitize_messages(msgs: list[Any]) -> list[Any]:
"""Keep only system/human/assistant messages and ensure content is str."""
clean = []
for m in msgs or []:
# Drop raw ToolMessage or unknown roles (Gemini doesn't accept them)
role = getattr(m, "type", None) or getattr(m, "role", None) or ""
if isinstance(m, ToolMessage) or role == "tool":
# Optionally compress tool outputs into a short assistant line instead:
txt = _to_text(getattr(m, "content", None))
if txt:
clean.append(AIMessage(content=f"[Tool result] {txt}"))
continue
c = _to_text(getattr(m, "content", None))
if isinstance(m, SystemMessage):
clean.append(SystemMessage(content=c))
elif isinstance(m, HumanMessage):
clean.append(HumanMessage(content=c))
elif isinstance(m, AIMessage):
clean.append(AIMessage(content=c))
else:
# Unknown BaseMessage; best-effort map by role string
r = str(role).lower()
if r == "system":
clean.append(SystemMessage(content=c))
elif r in ("human", "user"):
clean.append(HumanMessage(content=c))
elif r in ("assistant", "ai", "aimessage"):
clean.append(AIMessage(content=c))
# else: ignore silently
return clean
TOOLS = [tool_inspect_dataset, tool_sota_preprocessing, tool_list_steps, tool_describe_step, tool_propose_plan, tool_run_step, tool_list_versions, tool_reset_to_version]
SYSTEM_PRIMER = (
"You are a data-quality assistant.\n"
"\n"
"Workflow:\n"
"1) Call inspect_dataset() to summarize columns/dtypes and GUESS task/label.\n"
" • If you are NOT SURE about the task (or the label for supervised tasks), ASK the user to confirm and END THE TURN.\n"
" • Do NOT call sota_preprocessing until the user explicitly confirms the task (and label if supervised).\n"
" Acceptable confirmations include messages like: "
" 'task=classification label=HARDSHIP_INDEX', 'Task: regression', or 'Unsupervised'.\n"
"2) After the user confirms, call sota_preprocessing(task, modality, ...) and PRESENT a brief 'SOTA Evidence' section (3–6 bullets with titles and links from the tool).\n"
"3) Call list_steps() and map SOTA insights to the available tools. Produce a plan (no execution yet); cite up to 2 SOTA sources per step.\n"
"4) Ask: 'Which step should we execute first?' Do NOT call run_step until the user explicitly picks.\n"
"5) After the user picks, call describe_step(name) and list ONLY real parameters from the tool. Ask for missing/optional params and confirm them.\n"
"6) Execute with run_step(name, params_json). Version controls inside params_json when relevant:\n"
" • source: 'current' | 'prev' | 'base' | '@-1' | '@-2' | <int>\n"
" • dry_run: true|false (preview without mutating)\n"
" • new_version: true|false (create new snapshot vs replace current)\n"
" Avoid loops: if the same step+params just ran, ask to change parameters or source.\n"
"7) Summarize results; optionally call list_versions() and offer reset_to_version(spec). If helpful, research again before proposing next steps.\n"
"\n"
"Rules:\n"
"- Return exactly one tool call at a time.\n"
"- Never call sota_preprocessing before explicit task confirmation.\n"
"- Never call run_step without an explicit user choice.\n"
"- When users ask about parameters, use describe_step (or list_steps) and answer ONLY from tool output.\n"
"- Reject parameters that are not in the tool signature.\n"
)
class AgentState(TypedDict):
messages: List[Any]
df_payload: Optional[Dict[str, Any]]
results: List[Dict[str, Any]]
steps_taken: int
max_steps: int
confirmed_step: Optional[str]
confirmed_params: Dict[str, Any]
last_task: Optional[str]
plan: Optional[Dict[str, Any]]
def make_agent_node(llm):
"""LLM emits tool calls; we sanitize history and ALWAYS append an AIMessage."""
llm_with_tools = llm.bind_tools(TOOLS)
def _node(state: AgentState) -> AgentState:
d = (state.get("df_payload") or {}).get("data", {})
rows = len(d.get("data", []) or [])
cols = len(d.get("columns", []) or [])
shape_note = SystemMessage(content=f"Current dataset shape: {rows} rows × {cols} columns.")
history = _sanitize_messages(state.get("messages", []))
inputs = [SystemMessage(content=SYSTEM_PRIMER), *history, shape_note]
ai = llm_with_tools.invoke(inputs)
# guard: ensure we append an AIMessage object
if not isinstance(ai, AIMessage):
ai = AIMessage(content=_to_text(getattr(ai, "content", ai)))
state["messages"] = state["messages"] + [ai]
# debug
# print("DEBUG roles after agent:", [getattr(m, "type", None) or getattr(m, "role", None) for m in state["messages"]])
return state
return _node
def tools_exec_node():
"""
Execute tools only here, after injecting df_payload into runtime context.
Also updates state with tool outputs (summary/SOTA/plan/step_result).
"""
tool_node = ToolNode(TOOLS)
def _node(state: AgentState) -> AgentState:
# Inject dataset into runtime context BEFORE any tool executes
set_df_payload(state.get("df_payload"))
# If no dataset at all, be friendly and stop
if state.get("df_payload") is None:
state["messages"].append(type(state["messages"][-1])(content="I don't have a dataset yet. Please upload one."))
return state
# Hard gate: block run_step unless user confirmed a step
last = state["messages"][-1]
tool_calls = getattr(last, "tool_calls", None) or []
for c in tool_calls:
if c.get("name") == "run_step":
intended = (c.get("args") or {}).get("name")
if intended and intended != state.get("confirmed_step"):
state["messages"].append(type(last)(content="I have a plan ready. Which step should we run first?"))
return state
# Actually execute the tool(s) requested by the last assistant message
out = tool_node.invoke({"messages": state["messages"]})
# Append ONLY new ToolMessages; do NOT overwrite the conversation
new_msgs = [m for m in out["messages"] if isinstance(m, ToolMessage)]
if not new_msgs:
# fallback: if provider returned the whole list, take the tail
if len(out["messages"]) > len(state["messages"]):
new_msgs = out["messages"][len(state["messages"]):]
else:
new_msgs = out["messages"]
state["messages"] = state["messages"] + new_msgs
# Parse the most recent tool payload (dict in .content)
payload = new_msgs[-1].content if new_msgs else None
if isinstance(payload, dict):
typ = payload.get("type")
if typ == "dataset_summary":
set_df_summary(payload)
state["last_task"] = payload.get("task_guess")
elif typ == "sota":
set_sota_bundled(payload.get("bundled_results") or [])
elif typ == "plan":
state["plan"] = payload
elif typ == "step_result":
state["df_payload"] = payload["df"]
set_df_payload(state["df_payload"])
state["results"].append({"name": payload["name"], "stats": payload["stats"]})
state["steps_taken"] += 1
state["confirmed_step"] = None
state["confirmed_params"] = {}
# print("DEBUG roles after tools:", [getattr(m, "type", None) or getattr(m, "role", None) for m in state["messages"]])
return state
return _node
def should_continue(state: AgentState) -> str:
last = state["messages"][-1]
if state.get("steps_taken", 0) >= state.get("max_steps", 8):
return "end"
# Continue if the last assistant message contains tool calls
return "continue" if getattr(last, "tool_calls", None) else "end"
def build_app(llm):
g = StateGraph(AgentState)
g.add_node("agent", make_agent_node(llm))
g.add_node("tools", tools_exec_node())
g.add_edge(START, "agent")
g.add_conditional_edges("agent", should_continue, {"continue": "tools", "end": END})
g.add_edge("tools", "agent")
return g.compile()
|