Eras Expert domain context: - Full Heizkostenabrechnung business model (Kunde>Objekte>Nutzeinheiten>Geraete) - Known PK/FK mappings: kunden.Kundennummer, objekte.KundenID, etc. - Correct JOIN example in SCHEMA prompt - PA knows domain hierarchy for better job formulation Iterative plan-execute in ExpertNode: - DESCRIBE queries execute first, results injected into re-plan - Re-plan uses actual column names from DESCRIBE - Eliminates "Unknown column" errors on first query Frontend: - Node inspector: per-node cards with model, tokens, progress, last event - Graph switcher buttons in top bar - Clear button in top bar - Nodes panel 300px wide - WS reconnect on 1006 (deploy) without showing login - Model info emitted on context HUD events Domain context test: 21/21 (hierarchy, JOINs, FK, PA job quality) Default graph: v4-eras Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
222 lines
8.6 KiB
Python
222 lines
8.6 KiB
Python
"""Expert Base Node: domain-specific stateless executor.
|
|
|
|
An expert receives a self-contained job from the PA, plans its own tool sequence,
|
|
executes tools, and returns a ThoughtResult. No history, no memory — pure function.
|
|
|
|
Subclasses override DOMAIN_SYSTEM, SCHEMA, and default_database.
|
|
"""
|
|
|
|
import asyncio
|
|
import json
|
|
import logging
|
|
|
|
from .base import Node
|
|
from ..llm import llm_call
|
|
from ..db import run_db_query
|
|
from ..types import ThoughtResult
|
|
|
|
log = logging.getLogger("runtime")
|
|
|
|
|
|
class ExpertNode(Node):
|
|
"""Base class for domain experts. Subclass and set DOMAIN_SYSTEM, SCHEMA, default_database."""
|
|
|
|
model = "google/gemini-2.0-flash-001"
|
|
max_context_tokens = 4000
|
|
|
|
# Override in subclasses
|
|
DOMAIN_SYSTEM = "You are a domain expert."
|
|
SCHEMA = ""
|
|
default_database = "eras2_production"
|
|
|
|
PLAN_SYSTEM = """You are a domain expert's planning module.
|
|
Given a job description, produce a JSON tool sequence to accomplish it.
|
|
|
|
{domain}
|
|
|
|
{schema}
|
|
|
|
Available tools:
|
|
- query_db(query, database) — SQL SELECT/DESCRIBE/SHOW only
|
|
- emit_actions(actions) — show buttons [{{label, action, payload?}}]
|
|
- set_state(key, value) — persistent key-value
|
|
- emit_display(items) — formatted data [{{type, label, value?, style?}}]
|
|
- create_machine(id, initial, states) — interactive UI with navigation
|
|
states: {{"state_name": {{"actions": [...], "display": [...]}}}}
|
|
- add_state(id, state, buttons, content) — add state to machine
|
|
- reset_machine(id) — reset to initial
|
|
- destroy_machine(id) — remove machine
|
|
|
|
Output ONLY valid JSON:
|
|
{{
|
|
"tool_sequence": [
|
|
{{"tool": "query_db", "args": {{"query": "SELECT ...", "database": "{database}"}}}},
|
|
{{"tool": "emit_actions", "args": {{"actions": [{{"label": "...", "action": "..."}}]}}}}
|
|
],
|
|
"response_hint": "How to phrase the result for the user"
|
|
}}
|
|
|
|
Rules:
|
|
- NEVER guess column names. If unsure, DESCRIBE first.
|
|
- Max 5 tools. Keep it focused.
|
|
- The job is self-contained — all context you need is in the job description."""
|
|
|
|
RESPONSE_SYSTEM = """You are a domain expert summarizing results for the user.
|
|
|
|
{domain}
|
|
|
|
Job: {job}
|
|
{results}
|
|
|
|
Write a concise, natural response. 1-3 sentences.
|
|
- Reference specific data from the results.
|
|
- Don't repeat raw output — summarize.
|
|
- Match the language: {language}."""
|
|
|
|
def __init__(self, send_hud, process_manager=None):
|
|
super().__init__(send_hud)
|
|
|
|
async def execute(self, job: str, language: str = "de") -> ThoughtResult:
|
|
"""Execute a self-contained job. Returns ThoughtResult.
|
|
Uses iterative plan-execute: if DESCRIBE queries are in the plan,
|
|
execute them first, inject results into a re-plan, then execute the rest."""
|
|
await self.hud("thinking", detail=f"planning: {job[:80]}")
|
|
|
|
# Step 1: Plan tool sequence
|
|
schema_context = self.SCHEMA
|
|
plan_messages = [
|
|
{"role": "system", "content": self.PLAN_SYSTEM.format(
|
|
domain=self.DOMAIN_SYSTEM, schema=schema_context,
|
|
database=self.default_database)},
|
|
{"role": "user", "content": f"Job: {job}"},
|
|
]
|
|
plan_raw = await llm_call(self.model, plan_messages)
|
|
tool_sequence, response_hint = self._parse_plan(plan_raw)
|
|
|
|
# Step 1b: Execute DESCRIBE queries first, then re-plan with actual schema
|
|
describe_results = {}
|
|
remaining_tools = []
|
|
for step in tool_sequence:
|
|
if step.get("tool") == "query_db":
|
|
query = step.get("args", {}).get("query", "").strip().upper()
|
|
if query.startswith("DESCRIBE") or query.startswith("SHOW"):
|
|
await self.hud("tool_call", tool="query_db", args=step.get("args", {}))
|
|
try:
|
|
result = await asyncio.to_thread(
|
|
run_db_query, step["args"]["query"],
|
|
step["args"].get("database", self.default_database))
|
|
describe_results[step["args"]["query"]] = result
|
|
await self.hud("tool_result", tool="query_db", output=result[:200])
|
|
except Exception as e:
|
|
await self.hud("tool_result", tool="query_db", output=str(e)[:200])
|
|
else:
|
|
remaining_tools.append(step)
|
|
else:
|
|
remaining_tools.append(step)
|
|
|
|
# Re-plan if we got DESCRIBE results (now we know actual column names)
|
|
if describe_results:
|
|
schema_update = "Actual column names from DESCRIBE:\n"
|
|
for q, result in describe_results.items():
|
|
schema_update += f"\n{q}:\n{result[:500]}\n"
|
|
|
|
replan_messages = [
|
|
{"role": "system", "content": self.PLAN_SYSTEM.format(
|
|
domain=self.DOMAIN_SYSTEM,
|
|
schema=schema_context + "\n\n" + schema_update,
|
|
database=self.default_database)},
|
|
{"role": "user", "content": f"Job: {job}\n\nUse ONLY the actual column names from DESCRIBE above. Do NOT include DESCRIBE steps — they are already done."},
|
|
]
|
|
replan_raw = await llm_call(self.model, replan_messages)
|
|
new_tools, new_hint = self._parse_plan(replan_raw)
|
|
if new_tools:
|
|
remaining_tools = new_tools
|
|
if new_hint:
|
|
response_hint = new_hint
|
|
|
|
tool_sequence = remaining_tools
|
|
await self.hud("planned", tools=len(tool_sequence), hint=response_hint[:80])
|
|
|
|
# Step 2: Execute remaining tools
|
|
actions = []
|
|
state_updates = {}
|
|
display_items = []
|
|
machine_ops = []
|
|
tool_used = ""
|
|
tool_output = ""
|
|
|
|
for step in tool_sequence:
|
|
tool = step.get("tool", "")
|
|
args = step.get("args", {})
|
|
await self.hud("tool_call", tool=tool, args=args)
|
|
|
|
if tool == "emit_actions":
|
|
actions.extend(args.get("actions", []))
|
|
elif tool == "set_state":
|
|
key = args.get("key", "")
|
|
if key:
|
|
state_updates[key] = args.get("value")
|
|
elif tool == "emit_display":
|
|
display_items.extend(args.get("items", []))
|
|
elif tool == "create_machine":
|
|
machine_ops.append({"op": "create", **args})
|
|
elif tool == "add_state":
|
|
machine_ops.append({"op": "add_state", **args})
|
|
elif tool == "reset_machine":
|
|
machine_ops.append({"op": "reset", **args})
|
|
elif tool == "destroy_machine":
|
|
machine_ops.append({"op": "destroy", **args})
|
|
elif tool == "query_db":
|
|
query = args.get("query", "")
|
|
database = args.get("database", self.default_database)
|
|
try:
|
|
result = await asyncio.to_thread(run_db_query, query, database)
|
|
tool_used = "query_db"
|
|
tool_output = result
|
|
await self.hud("tool_result", tool="query_db", output=result[:200])
|
|
except Exception as e:
|
|
tool_used = "query_db"
|
|
tool_output = f"Error: {e}"
|
|
await self.hud("tool_result", tool="query_db", output=str(e)[:200])
|
|
|
|
# Step 3: Generate response
|
|
results_text = ""
|
|
if tool_output:
|
|
results_text = f"Tool result:\n{tool_output[:500]}"
|
|
|
|
resp_messages = [
|
|
{"role": "system", "content": self.RESPONSE_SYSTEM.format(
|
|
domain=self.DOMAIN_SYSTEM, job=job, results=results_text, language=language)},
|
|
{"role": "user", "content": job},
|
|
]
|
|
response = await llm_call(self.model, resp_messages)
|
|
if not response:
|
|
response = "[no response]"
|
|
|
|
await self.hud("done", response=response[:100])
|
|
|
|
return ThoughtResult(
|
|
response=response,
|
|
tool_used=tool_used,
|
|
tool_output=tool_output,
|
|
actions=actions,
|
|
state_updates=state_updates,
|
|
display_items=display_items,
|
|
machine_ops=machine_ops,
|
|
)
|
|
|
|
def _parse_plan(self, raw: str) -> tuple[list, str]:
|
|
"""Parse tool sequence JSON from planning LLM call."""
|
|
text = raw.strip()
|
|
if text.startswith("```"):
|
|
text = text.split("\n", 1)[1] if "\n" in text else text[3:]
|
|
if text.endswith("```"):
|
|
text = text[:-3]
|
|
text = text.strip()
|
|
try:
|
|
data = json.loads(text)
|
|
return data.get("tool_sequence", []), data.get("response_hint", "")
|
|
except (json.JSONDecodeError, Exception) as e:
|
|
log.error(f"[expert] plan parse failed: {e}, raw: {text[:200]}")
|
|
return [], ""
|