toolforge-env / server /plan_evaluator.py
Hugging-coder14's picture
Upload folder using huggingface_hub
46b3240 verified
Raw
History Blame Contribute Delete
18.1 kB
"""
Plan Evaluator
Consolidates the core evaluation logic:
1. Structural validation
2. Semantic slot judgment
3. Reward calculation
4. Token-cost calculation
5. Final score combination
Functions are pure and deterministic where possible.
"""
import logging
import json
import os
from typing import Any, Dict, List, Optional, Tuple
from openai import OpenAI
from models import (
MacroProposal,
SlotJudgmentResult,
Task,
Tool,
ToolCall,
ToolEvaluation,
ValidationResult,
)
from server.llm_eval_prompts import (
SLOT_JUDGE_SYSTEM_PROMPT,
build_slot_judge_user_prompt,
)
from server.slots import DEVOPS_SLOTS
logger = logging.getLogger(__name__)
# --- Named Constants (New Bounded Reward Design) ---
# Final reward bounds
FINAL_REWARD_MIN = -0.2
FINAL_REWARD_MAX = 1.0
# Stage 1: Validation
VALIDATION_PENALTY = -0.2
# Stage 2: Slot score bounds
SLOT_THRESHOLD = 0.65
SLOT_SCORE_MIN = -0.15 # slot_ratio == 0.0
SLOT_SCORE_MAX = 0.25 # slot_ratio == 1.0
# Stage 3: Macro bonuses
MACRO_CREATION_MAX = 0.20
MACRO_CREATION_DECAY_FLOOR = 0.05
MACRO_CREATION_THRESHOLD = 2
MACRO_CREATION_FULL_RANGE = {2, 3} # counts that get full reward
MACRO_USAGE_PARTIAL = 0.03 # when 0.65 <= slot_ratio < 1.0
MACRO_USAGE_FULL = 0.05 # when slot_ratio == 1.0
# Stage 4: Tool efficiency bounds
EFFICIENCY_SCORE_BASELINE = 0.2 # exact baseline match
EFFICIENCY_SCORE_MIN = 0.0
EFFICIENCY_SCORE_MAX = 0.5
EFFICIENCY_SCALE = 0.3 # multiplier on efficiency_ratio
# LLM configuration
API_BASE_URL = os.getenv("API_BASE_URL", "https://router.huggingface.co/v1")
MODEL_NAME = os.getenv("MODEL_NAME", "Qwen/Qwen2.5-72B-Instruct")
HF_TOKEN = os.getenv("HF_TOKEN", "")
# --- Helper Functions for Stage 2 (Semantic Judge) ---
def _build_judge_request(
task_prompt: str,
required_slots: List[str],
slot_definitions: Dict[str, str],
available_tools: List[Tool],
plan: List[ToolCall],
) -> Dict[str, Any]:
"""Helper to build the expected LLM judge prompt/input."""
# Placeholder structure for when real call is integrated
return {
"task_prompt": task_prompt,
"required_slots": required_slots,
"slot_definitions": slot_definitions,
"tools": [t.name for t in available_tools],
"plan": [{"tool": c.tool_name} for c in plan]
}
def _simulate_llm_judgment(
judge_request: Dict[str, Any],
plan: List[ToolCall],
required_slots: List[str]
) -> List[Dict[str, Any]]:
"""Helper to simulate the LLM response deterministically.
Produces classification: 'relevant', 'unnecessary', or 'harmful'.
"""
results = []
slots_filled_so_far = set()
# We simulate semantic relevance by simply mapping the sequence
# to the required slots.
for i, call in enumerate(plan):
# Extremely naive heuristic for simulation:
if call.tool_name == "delete" or "drop" in call.tool_name:
# Simulate a harmful destructive call
classification = "harmful"
slot = None
else:
# If we haven't filled all slots and it's not a duplicate, let's pretend it fills a slot
if len(slots_filled_so_far) < len(required_slots):
slot = required_slots[len(slots_filled_so_far)]
classification = "relevant"
slots_filled_so_far.add(slot)
else:
slot = None
classification = "unnecessary"
results.append({
"tool_call_index": i,
"tool_name": call.tool_name,
"fills_slot": slot,
"classification": classification,
"reason": f"Simulated classification: {classification}"
})
return results
def _call_llm_slot_judgment(
task_prompt: str,
required_slots: List[str],
slot_definitions: Dict[str, str],
available_tools: List[Tool],
plan: List[ToolCall],
) -> Dict[str, Any]:
"""Call the OpenAI-compatible LLM and return its JSON response."""
client = OpenAI(base_url=API_BASE_URL, api_key=HF_TOKEN or None)
# Map tool names to descriptions for the prompt
tool_desc_map = {t.name: t.description for t in available_tools}
plan_with_descs = []
for call in plan:
desc = tool_desc_map.get(call.tool_name, "No description available.")
plan_with_descs.append({
"tool_name": call.tool_name,
"tool_description": desc
})
user_prompt = build_slot_judge_user_prompt(
task_prompt=task_prompt,
required_slots=required_slots,
slot_definitions=slot_definitions,
plan=plan_with_descs,
)
completion = client.chat.completions.create(
model=MODEL_NAME,
messages=[
{"role": "system", "content": SLOT_JUDGE_SYSTEM_PROMPT},
{"role": "user", "content": user_prompt},
],
temperature=0.0,
response_format={"type": "json_object"},
)
content = completion.choices[0].message.content or "{}"
content = content.strip().replace("```json", "").replace("```", "").strip()
return json.loads(content)
def _parse_llm_judgment(raw_json: Dict[str, Any], required_slots: List[str]) -> SlotJudgmentResult:
"""Helper to convert the flat LLM summary output into a SlotJudgmentResult.
The expected LLM schema is:
{
"slots_filled": ["SLOT_NAME"],
"slots_missing": ["SLOT_NAME"],
"unnecessary_calls": ["tool_name"],
"harmful_calls": ["tool_name"]
}
"""
slots_filled = raw_json.get("slots_filled", [])
slots_missing = raw_json.get("slots_missing", [])
# Validation against the requested required_slots
# Ensure slots_filled only contains requested slots
slots_filled = [s for s in slots_filled if s in required_slots]
# If missing is not explicitly provided, we compute it
if not slots_missing:
slots_missing = [s for s in required_slots if s not in slots_filled]
harmful_calls = raw_json.get("harmful_calls", [])
harmful_calls_present = len(harmful_calls) > 0
task_complete = len(slots_missing) == 0
# Create minimal ToolEvaluation entries if we have names, but since we lost the
# tool_call_index in the summary, we maintain an empty list for the model requirement.
# The current reward logic only depends on individual slot filling counts.
evaluations = []
return SlotJudgmentResult(
evaluations=evaluations,
slots_filled=slots_filled,
slots_missing=slots_missing,
task_complete=task_complete,
harmful_calls_present=harmful_calls_present,
)
def _simulate_llm_judgment(
judge_request: Dict[str, Any],
plan: List[ToolCall],
required_slots: List[str]
) -> Dict[str, Any]:
"""Helper to simulate the LLM response in the flat summary format.
Returns the same schema as defined in SLOT_JUDGE_SYSTEM_PROMPT.
"""
slots_filled = []
harmful_calls = []
unnecessary_calls = []
slots_filled_so_far = set()
for call in plan:
# Heuristic for simulation
if "delete" in call.tool_name or "drop" in call.tool_name:
harmful_calls.append(call.tool_name)
elif len(slots_filled_so_far) < len(required_slots):
slot = required_slots[len(slots_filled_so_far)]
slots_filled.append(slot)
slots_filled_so_far.add(slot)
else:
unnecessary_calls.append(call.tool_name)
return {
"slots_filled": slots_filled,
"slots_missing": [s for s in required_slots if s not in slots_filled],
"unnecessary_calls": unnecessary_calls,
"harmful_calls": harmful_calls
}
# --- Helper Function for Stage 4 (Dynamic Baseline) ---
def calculate_dynamic_baseline_tokens(task: Task, available_tools: Dict[str, Tool]) -> int:
"""Computes the expected baseline token cost for a task.
Slot-driven baseline: uses baseline_token_cost when provided,
otherwise falls back to baseline_call_count for compatibility.
"""
if task.baseline_call_count > 0:
return task.baseline_call_count
return 0
# --- Helper Functions for Macro Recognition ---
def extract_contiguous_windows(tool_names: List[str], window_size: int) -> List[Tuple[str, ...]]:
"""Return exact ordered contiguous windows of the given size."""
if window_size < 2 or window_size > len(tool_names):
return []
return [tuple(tool_names[i:i + window_size]) for i in range(len(tool_names) - window_size + 1)]
def count_prior_sequence_occurrences(
proposed_sequence: Tuple[str, ...],
sequence_counts: Dict[str, int],
) -> int:
"""Return the prior exact count for a sequence key."""
key = str(proposed_sequence)
return sequence_counts.get(key, 0)
def update_sequence_counts(
plan: List[ToolCall],
sequence_counts: Dict[str, int],
) -> Dict[str, int]:
"""Update sequence_counts dict from the current plan's contiguous windows.
Extracts windows of size 2..len(plan) and increments counts.
Returns the updated dict (mutates in place for convenience).
"""
tool_names = [call.tool_name for call in plan]
for window_size in range(2, len(tool_names) + 1):
for window in extract_contiguous_windows(tool_names, window_size):
key = str(window)
sequence_counts[key] = sequence_counts.get(key, 0) + 1
return sequence_counts
# --- New Bounded Stage Scoring Functions ---
def compute_slot_score(slot_ratio: float) -> float:
"""Stage 2: Piecewise linear slot score bounded to [-0.15, 0.25].
- slot_ratio < 0.65: maps [-0.15, 0.0]
- slot_ratio >= 0.65: maps [0.0, 0.25]
"""
if slot_ratio < SLOT_THRESHOLD:
# linear from -0.15 (at 0.0) to 0.0 (at 0.65)
return SLOT_SCORE_MIN * (1.0 - slot_ratio / SLOT_THRESHOLD)
else:
# linear from 0.0 (at 0.65) to 0.25 (at 1.0)
return SLOT_SCORE_MAX * ((slot_ratio - SLOT_THRESHOLD) / (1.0 - SLOT_THRESHOLD))
def compute_macro_creation_bonus(
macro_proposal: Optional[Tool],
sequence_counts: Optional[Dict[str, int]],
slot_ratio: float,
) -> float:
"""Stage 3a: Macro creation bonus bounded to [0.0, 0.20].
Gate: slot_ratio >= 0.65
- prior_count < 2: 0.0
- prior_count in {2, 3}: 0.20
- prior_count > 3: decays with floor 0.05
"""
if slot_ratio < SLOT_THRESHOLD:
return 0.0
if macro_proposal is None or sequence_counts is None:
return 0.0
if macro_proposal.steps is None or len(macro_proposal.steps) < 2:
return 0.0
proposed_sequence = tuple(call.tool_name for call in macro_proposal.steps)
prior_count = count_prior_sequence_occurrences(proposed_sequence, sequence_counts)
if prior_count < MACRO_CREATION_THRESHOLD:
return 0.0
if prior_count in MACRO_CREATION_FULL_RANGE:
return MACRO_CREATION_MAX
# Decay for late creation
return max(MACRO_CREATION_DECAY_FLOOR, MACRO_CREATION_MAX * (3.0 / prior_count))
def compute_macro_usage_bonus(
plan: List[ToolCall],
accepted_macros: List[Tool],
slot_ratio: float,
) -> float:
"""Stage 3b: Macro usage bonus bounded to [0.0, 0.05].
Gate: slot_ratio >= 0.65
"""
if slot_ratio < SLOT_THRESHOLD:
return 0.0
macro_names = {m.name for m in accepted_macros}
macro_used = any(call.tool_name in macro_names for call in plan)
if not macro_used:
return 0.0
if slot_ratio >= 1.0:
return MACRO_USAGE_FULL
return MACRO_USAGE_PARTIAL
def compute_efficiency_score(
plan: List[ToolCall],
task: Task,
available_tools: Dict[str, Tool],
) -> float:
"""Stage 4: Count-based efficiency score bounded to [0.0, 0.5].
Only called when slot_ratio == 1.0.
- baseline match -> 0.2
- better than baseline -> above 0.2
- worse than baseline -> below 0.2
"""
baseline = calculate_dynamic_baseline_tokens(task, available_tools)
actual = len(plan)
if baseline <= 0:
return EFFICIENCY_SCORE_BASELINE
efficiency_ratio = (baseline - actual) / baseline
score = EFFICIENCY_SCORE_BASELINE + EFFICIENCY_SCALE * efficiency_ratio
return max(EFFICIENCY_SCORE_MIN, min(EFFICIENCY_SCORE_MAX, score))
# --- Public APIs ---
def get_relevant_slots(required_slots: List[str]) -> Dict[str, str]:
"""Return the subset of DEVOPS_SLOTS matching required_slots."""
return {
name: DEVOPS_SLOTS[name]
for name in required_slots
if name in DEVOPS_SLOTS
}
def run_sanity_validation(
plan: List[ToolCall],
available_tools: Dict[str, Tool],
) -> ValidationResult:
"""Stage 1: Structural validation of the plan.
Validation order:
1. Empty plan → EMPTY_PLAN (penalty -0.2)
2. Unknown tool name → INVALID_TOOL (penalty -0.2)
3. All pass → VALID (penalty 0.0)
"""
if plan is None or len(plan) == 0:
return ValidationResult(
valid=False,
reason="EMPTY_PLAN",
penalty=VALIDATION_PENALTY,
detail="Plan contains no tool calls.",
)
for call in plan:
if call.tool_name not in available_tools:
return ValidationResult(
valid=False,
reason="INVALID_TOOL",
penalty=VALIDATION_PENALTY,
detail=f"Tool '{call.tool_name}' does not exist in toolbox.",
)
return ValidationResult(valid=True, reason="VALID", penalty=0.0)
def _expand_macros_in_plan(plan: List[ToolCall], available_tools: List[Tool]) -> List[ToolCall]:
"""Recursively expand macro calls into their atomic components for semantic evaluation."""
tool_map = {t.name: t for t in available_tools}
expanded_plan = []
for call in plan:
tool = tool_map.get(call.tool_name)
if tool and tool.is_macro and tool.steps:
# Macros are typically 1-level, but we expand recursively for robustness
expanded_plan.extend(_expand_macros_in_plan(tool.steps, available_tools))
else:
expanded_plan.append(call)
return expanded_plan
def run_slot_judgment(
task_prompt: str,
required_slots: List[str],
slot_definitions: Dict[str, str],
available_tools: List[Tool],
plan: List[ToolCall],
) -> SlotJudgmentResult:
"""Stage 2: Evaluate a validated plan against the task's semantic slots."""
# Expand macros into atomic steps so the LLM evaluator can use tool descriptions
# expanded_plan = _expand_macros_in_plan(plan, available_tools)
# logger.debug(f"Stage 2 expansion: original_len={len(plan)}, expanded_len={len(expanded_plan)}")
req = _build_judge_request(task_prompt, required_slots, slot_definitions, available_tools, plan)
try:
raw_json = _call_llm_slot_judgment(
task_prompt=task_prompt,
required_slots=required_slots,
slot_definitions=slot_definitions,
available_tools=available_tools,
plan=plan,
)
if not raw_json or "slots_filled" not in raw_json:
raise ValueError("LLM returned an invalid or empty response")
except Exception as exc:
logger.warning("LLM slot judge failed, falling back to simulated judgment: %s", exc)
raw_json = _simulate_llm_judgment(req, plan, required_slots)
result = _parse_llm_judgment(raw_json, required_slots)
if result.harmful_calls_present:
logger.warning("Slot judge detected harmful calls in plan.")
return result
def compute_step_reward(
slot_judgment: SlotJudgmentResult,
task: Task,
plan: List[ToolCall],
available_tools: Dict[str, Tool],
accepted_macros: List[Tool],
macro_proposal: Optional[Tool] = None,
sequence_counts: Optional[Dict[str, int]] = None,
) -> Dict[str, float]:
"""Compute the full step reward using bounded additive stages.
Returns a dict with per-stage contributions and the final clamped reward.
"""
# Compute slot ratio
n_required = len(task.required_slots)
n_filled = len(slot_judgment.slots_filled)
slot_ratio = n_filled / n_required if n_required > 0 else 1.0
# Stage 2: slot score
slot_score = compute_slot_score(slot_ratio)
# Stage 3: macro bonuses (gated by slot_ratio)
macro_creation = compute_macro_creation_bonus(macro_proposal, sequence_counts, slot_ratio)
macro_usage = compute_macro_usage_bonus(plan, accepted_macros, slot_ratio)
# Stage 4: efficiency (only when fully complete)
efficiency_score = 0.0
if slot_ratio < SLOT_THRESHOLD:
# Only slot score, no macro, no efficiency
final_raw = slot_score
elif slot_ratio < 1.0:
# Slot score + macro bonuses only
final_raw = slot_score + macro_creation + macro_usage
else:
# Full: slot + macro + efficiency
efficiency_score = compute_efficiency_score(plan, task, available_tools)
final_raw = slot_score + macro_creation + macro_usage + efficiency_score
final_reward = max(FINAL_REWARD_MIN, min(FINAL_REWARD_MAX, final_raw))
logger.info("Final Reward: %.3f", final_reward)
return {
"slot_ratio": slot_ratio,
"slot_score": slot_score,
"macro_creation": macro_creation,
"macro_usage": macro_usage,
"efficiency_score": efficiency_score,
"final_reward": final_reward,
}