d3f7099d93
- Rewrite eval config with EvalMode (step1, step2, full) replacing prompt_variants - Rewrite runner with _run_step1, _run_step2, _run_full dispatch - CLI: replace --variants with --mode flag - Add 3 fixture YAMLs: classify_invoices (step1), process_invoices (step2), full_invoices (full) - Remove old freelance_invoices fixture - Langfuse: mode-aware dataset items (classifications for step1, extraction for step2, both for full) - Langfuse: link both prompts (batch_file_classifier + batch_processing) in full mode - Langfuse: post separate classification_precision/recall/f1 scores for full mode - Langfuse: skip misleading field_accuracy=0 when field_scores is empty (step1) - Langfuse: include step1_results in trace output - MockExecutor: mock async_session to bypass DB in full mode - Journey fixture: remove user_messages (only interactive test kept)
386 lines
13 KiB
Python
386 lines
13 KiB
Python
"""Journey eval runner — tests the prompt_template builder conversation.
|
||
|
||
For each (journey_fixture × model) combination:
|
||
1. Build a MockExecutor (for filesystem tools used during journey)
|
||
2. Patch execute_on_client
|
||
3. Override LLM_MODEL
|
||
4. Call handle_journey_start to kick off the conversation
|
||
5. Feed simulated user_messages via handle_journey_message
|
||
6. Collect the generated prompt_template
|
||
7. Score it against expected_template_criteria (via LLM judge)
|
||
8. Report to Langfuse
|
||
"""
|
||
|
||
from __future__ import annotations
|
||
|
||
import asyncio
|
||
import copy
|
||
import json
|
||
import logging
|
||
import time
|
||
import uuid
|
||
from dataclasses import dataclass, field
|
||
from pathlib import Path
|
||
from typing import Any
|
||
|
||
from langchain_core.messages import HumanMessage, SystemMessage
|
||
|
||
from eval.config import JourneyFixture
|
||
from eval.mock_executor import MockExecutor
|
||
from eval import langfuse_eval
|
||
|
||
logger = logging.getLogger(__name__)
|
||
|
||
|
||
# ── Result type ──────────────────────────────────────────────────────────
|
||
|
||
|
||
@dataclass
|
||
class JourneyEvalResult:
|
||
"""Result of one journey eval run."""
|
||
|
||
fixture_name: str
|
||
model: str
|
||
prompt_template: str | None # the generated template (None if journey failed)
|
||
conversation_turns: int
|
||
done: bool # whether journey reached completion
|
||
criteria_scores: dict[str, float] # criterion → 0-1 score
|
||
overall_score: float # average of criteria scores
|
||
judge_reasoning: str
|
||
elapsed_seconds: float
|
||
|
||
def summary(self) -> dict[str, Any]:
|
||
return {
|
||
"fixture": self.fixture_name,
|
||
"model": self.model,
|
||
"done": self.done,
|
||
"turns": self.conversation_turns,
|
||
"overall_score": round(self.overall_score, 3),
|
||
"criteria_scores": {k: round(v, 3) for k, v in self.criteria_scores.items()},
|
||
"elapsed_s": round(self.elapsed_seconds, 1),
|
||
}
|
||
|
||
|
||
# ── LLM judge for template quality ──────────────────────────────────────
|
||
|
||
_JOURNEY_JUDGE_SYSTEM = """\
|
||
You are an evaluation judge for AI-generated prompt templates.
|
||
|
||
A journey chatbot explored a user's directory structure and through
|
||
conversation produced a prompt_template — an instruction set for a
|
||
data-extraction agent.
|
||
|
||
Your task: evaluate the generated template against a list of criteria.
|
||
Score each criterion from 0 to 1:
|
||
- 1.0: Fully satisfied, clearly present in the template
|
||
- 0.5: Partially satisfied or ambiguously addressed
|
||
- 0.0: Not satisfied, missing from the template
|
||
|
||
Respond with ONLY a JSON object:
|
||
{
|
||
"scores": {"criterion_1": 0.8, "criterion_2": 1.0, ...},
|
||
"reasoning": "Brief explanation"
|
||
}
|
||
"""
|
||
|
||
|
||
async def _judge_template(
|
||
prompt_template: str,
|
||
criteria: list[str],
|
||
*,
|
||
judge_model: str = "gpt-4o-mini",
|
||
) -> tuple[dict[str, float], str]:
|
||
"""Use an LLM to evaluate a generated prompt_template against criteria.
|
||
|
||
Returns (criteria_scores, reasoning).
|
||
"""
|
||
from shared.llm import get_llm
|
||
|
||
llm = get_llm(model=judge_model, temperature=0)
|
||
|
||
criteria_text = "\n".join(f" {i+1}. {c}" for i, c in enumerate(criteria))
|
||
user_content = (
|
||
f"## Generated prompt_template\n```\n{prompt_template}\n```\n\n"
|
||
f"## Criteria to evaluate\n{criteria_text}"
|
||
)
|
||
|
||
try:
|
||
response = await llm.ainvoke([
|
||
SystemMessage(content=_JOURNEY_JUDGE_SYSTEM),
|
||
HumanMessage(content=user_content),
|
||
])
|
||
raw = response.content.strip()
|
||
if raw.startswith("```"):
|
||
raw = raw.split("```")[1]
|
||
if raw.startswith("json"):
|
||
raw = raw[4:]
|
||
parsed = json.loads(raw.strip())
|
||
|
||
scores_raw = parsed.get("scores", {})
|
||
# Map criterion keys back to the original criteria text
|
||
criteria_scores: dict[str, float] = {}
|
||
for i, criterion in enumerate(criteria):
|
||
# Try matching by index key or exact criterion text
|
||
key_candidates = [
|
||
f"criterion_{i+1}",
|
||
criterion,
|
||
criterion[:50],
|
||
str(i + 1),
|
||
]
|
||
score = 0.0
|
||
for key in key_candidates:
|
||
if key in scores_raw:
|
||
score = float(scores_raw[key])
|
||
break
|
||
# If no match found, try values in order
|
||
if score == 0.0 and i < len(scores_raw):
|
||
score = float(list(scores_raw.values())[i])
|
||
criteria_scores[criterion] = score
|
||
|
||
reasoning = str(parsed.get("reasoning", ""))
|
||
return criteria_scores, reasoning
|
||
except Exception as exc:
|
||
logger.warning("journey_eval: LLM judge failed: %s", exc)
|
||
return {c: 0.0 for c in criteria}, f"Judge error: {exc}"
|
||
|
||
|
||
# ── Journey runner ───────────────────────────────────────────────────────
|
||
|
||
|
||
async def run_single_journey_eval(
|
||
fixture: JourneyFixture,
|
||
model: str,
|
||
*,
|
||
judge_model: str = "gpt-4o-mini",
|
||
data_dir: Path | None = None,
|
||
) -> JourneyEvalResult:
|
||
"""Execute one journey eval: start \u2192 messages \u2192 score template."""
|
||
from shared.config import settings
|
||
|
||
# When data_dir is given, use its parent as MockExecutor root
|
||
# and its name as the journey directory so the LLM sees a
|
||
# meaningful path (not ".").
|
||
if data_dir:
|
||
mock_root = data_dir.parent
|
||
journey_directory = data_dir.name
|
||
else:
|
||
mock_root = fixture.fixture_path.parent
|
||
journey_directory = fixture.directory
|
||
|
||
mock = MockExecutor(
|
||
fixture_dir=mock_root,
|
||
seed_records={},
|
||
)
|
||
|
||
original_model = settings.LLM_MODEL
|
||
settings.LLM_MODEL = model
|
||
|
||
eval_user_id = f"eval-journey-{uuid.uuid4().hex[:8]}"
|
||
|
||
logger.info(
|
||
"journey_eval: starting %s | model=%s",
|
||
fixture.name, model,
|
||
)
|
||
start_time = time.time()
|
||
|
||
prompt_template: str | None = None
|
||
conversation: list[dict[str, str]] = []
|
||
done = False
|
||
|
||
try:
|
||
from shared.ws_context import set_current_user, clear_current_user
|
||
from app.journey import handle_journey_start, handle_journey_message, _sessions
|
||
|
||
set_current_user(eval_user_id)
|
||
with mock.patch():
|
||
# ── Start the journey ────────────────────────────────
|
||
start_frame: dict[str, Any] = {
|
||
"agent_type": "local",
|
||
"directory": journey_directory,
|
||
"data_types": fixture.data_types,
|
||
"session_id": f"eval-{uuid.uuid4().hex[:8]}",
|
||
}
|
||
|
||
reply = await handle_journey_start(eval_user_id, start_frame)
|
||
session_id = reply["session_id"]
|
||
conversation.append({"role": "assistant", "content": reply["message"]})
|
||
|
||
logger.info(
|
||
"journey_eval: start reply (%d chars), done=%s",
|
||
len(reply["message"]), reply["done"],
|
||
)
|
||
|
||
if reply["done"]:
|
||
prompt_template = reply.get("prompt_template")
|
||
done = True
|
||
else:
|
||
# ── Send user messages ───────────────────────────
|
||
for i, user_msg in enumerate(fixture.user_messages):
|
||
if done:
|
||
break
|
||
|
||
conversation.append({"role": "user", "content": user_msg})
|
||
|
||
msg_frame: dict[str, Any] = {
|
||
"session_id": session_id,
|
||
"message": user_msg,
|
||
}
|
||
reply = await handle_journey_message(eval_user_id, msg_frame)
|
||
conversation.append({"role": "assistant", "content": reply["message"]})
|
||
|
||
logger.info(
|
||
"journey_eval: turn %d reply (%d chars), done=%s",
|
||
i + 1, len(reply["message"]), reply["done"],
|
||
)
|
||
|
||
if reply["done"]:
|
||
prompt_template = reply.get("prompt_template")
|
||
done = True
|
||
|
||
# If not done after all user messages, send a final nudge
|
||
if not done:
|
||
nudge = "Please generate the final prompt_template now. I'm satisfied with the configuration."
|
||
conversation.append({"role": "user", "content": nudge})
|
||
|
||
nudge_frame: dict[str, Any] = {
|
||
"session_id": session_id,
|
||
"message": nudge,
|
||
}
|
||
reply = await handle_journey_message(eval_user_id, nudge_frame)
|
||
conversation.append({"role": "assistant", "content": reply["message"]})
|
||
if reply["done"]:
|
||
prompt_template = reply.get("prompt_template")
|
||
done = True
|
||
|
||
except Exception as exc:
|
||
logger.error("journey_eval: pipeline failed for %s/%s: %s", fixture.name, model, exc)
|
||
finally:
|
||
settings.LLM_MODEL = original_model
|
||
from shared.ws_context import clear_current_user
|
||
clear_current_user()
|
||
|
||
elapsed = time.time() - start_time
|
||
turns = len([c for c in conversation if c["role"] == "user"])
|
||
|
||
logger.info(
|
||
"journey_eval: completed in %.1fs — %d turns, done=%s, template=%s",
|
||
elapsed, turns, done, "yes" if prompt_template else "no",
|
||
)
|
||
|
||
# ── Score the template ───────────────────────────────────────
|
||
criteria_scores: dict[str, float] = {}
|
||
judge_reasoning = ""
|
||
|
||
if prompt_template and fixture.expected_template_criteria:
|
||
criteria_scores, judge_reasoning = await _judge_template(
|
||
prompt_template,
|
||
fixture.expected_template_criteria,
|
||
judge_model=judge_model,
|
||
)
|
||
elif not prompt_template:
|
||
criteria_scores = {c: 0.0 for c in fixture.expected_template_criteria}
|
||
judge_reasoning = "No prompt_template was generated — journey did not complete."
|
||
|
||
overall = (
|
||
sum(criteria_scores.values()) / len(criteria_scores)
|
||
if criteria_scores
|
||
else 0.0
|
||
)
|
||
|
||
result = JourneyEvalResult(
|
||
fixture_name=fixture.name,
|
||
model=model,
|
||
prompt_template=prompt_template,
|
||
conversation_turns=turns,
|
||
done=done,
|
||
criteria_scores=criteria_scores,
|
||
overall_score=overall,
|
||
judge_reasoning=judge_reasoning,
|
||
elapsed_seconds=elapsed,
|
||
)
|
||
|
||
# ── Report to Langfuse ───────────────────────────────────────
|
||
trace_id = langfuse_eval.log_eval_trace(
|
||
fixture_name=fixture.name,
|
||
model=model,
|
||
prompt_variant="journey",
|
||
prompt_template=prompt_template or "(not generated)",
|
||
actual_mutations=[{"conversation": conversation[:20]}],
|
||
scores_summary=result.summary(),
|
||
langfuse_prompt_names=["journey_system"],
|
||
)
|
||
|
||
if trace_id:
|
||
from eval.scorer import EvalScores
|
||
scores_obj = EvalScores(
|
||
fixture_name=fixture.name,
|
||
model=model,
|
||
prompt_variant="journey",
|
||
precision=overall,
|
||
recall=float(done),
|
||
f1=overall,
|
||
llm_judge_score=overall,
|
||
llm_judge_reasoning=judge_reasoning,
|
||
)
|
||
langfuse_eval.post_eval_scores(scores_obj, trace_id=trace_id)
|
||
|
||
return result
|
||
|
||
|
||
async def run_journey_fixture_eval(
|
||
fixture: JourneyFixture,
|
||
models: list[str],
|
||
*,
|
||
judge_model: str = "gpt-4o-mini",
|
||
data_dir: Path | None = None,
|
||
) -> list[JourneyEvalResult]:
|
||
"""Run all models for a journey fixture."""
|
||
langfuse_eval.sync_journey_fixture_to_dataset(fixture)
|
||
|
||
results: list[JourneyEvalResult] = []
|
||
for model in models:
|
||
result = await run_single_journey_eval(
|
||
fixture, model, judge_model=judge_model,
|
||
data_dir=data_dir,
|
||
)
|
||
results.append(result)
|
||
|
||
return results
|
||
|
||
|
||
def print_journey_results(results: list[JourneyEvalResult]) -> None:
|
||
"""Print a formatted summary of journey eval results."""
|
||
if not results:
|
||
print("\nNo journey eval results.")
|
||
return
|
||
|
||
print("\n" + "=" * 95)
|
||
print(f"{'Fixture':<25} {'Model':<25} {'Done':>5} {'Turns':>6} {'Score':>7} {'Time':>7}")
|
||
print("-" * 95)
|
||
|
||
for r in results:
|
||
done_str = "yes" if r.done else "NO"
|
||
print(
|
||
f"{r.fixture_name:<25} {r.model:<25} {done_str:>5} "
|
||
f"{r.conversation_turns:>6} {r.overall_score:>7.2f} {r.elapsed_seconds:>6.1f}s"
|
||
)
|
||
|
||
print("=" * 95)
|
||
|
||
# Criteria breakdown
|
||
for r in results:
|
||
if r.criteria_scores:
|
||
print(f"\n[{r.model}] Criteria scores:")
|
||
for criterion, score in r.criteria_scores.items():
|
||
indicator = "PASS" if score >= 0.7 else "PARTIAL" if score >= 0.4 else "FAIL"
|
||
print(f" {indicator:>7} ({score:.1f}) {criterion}")
|
||
|
||
if r.judge_reasoning:
|
||
print(f" Judge: {r.judge_reasoning}")
|
||
|
||
if r.prompt_template:
|
||
preview = r.prompt_template[:200].replace("\n", " ")
|
||
print(f" Template preview: {preview}...")
|
||
|
||
print()
|