api / src /chatbot /guardrails.py
safraeli's picture
Vectorize Farquhar, DI ControlLoop, gate pipeline, budget audit, chatbot Hebrew
15be6bb verified
"""
Guardrails for the Vineyard Advisor chatbot.
Three components:
1. QueryClassifier โ€” determines if a query requires tool data or can be
answered from biology rules alone.
2. ResponseValidator โ€” deterministic post-response check that catches
rule violations before the answer reaches the user.
3. confidence_from_context โ€” estimates answer confidence based on data
freshness and availability.
"""
from __future__ import annotations
import re
from dataclasses import dataclass, field
from datetime import datetime
from typing import Optional
from config.settings import (
NO_SHADE_BEFORE_HOUR,
NO_SHADE_MONTHS,
NO_SHADE_TLEAF_BELOW,
)
# ---------------------------------------------------------------------------
# 1. Query classifier โ€” decides whether a tool call is mandatory
# ---------------------------------------------------------------------------
# Keywords that indicate user is asking about real-time / site-specific data
_DATA_KEYWORDS = [
# Weather / environment
r"\btemperature\b", r"\btemp\b", r"\bhow hot\b", r"\bhow cold\b",
r"\bweather\b", r"\bforecast\b", r"\brain\b", r"\bwind\b",
r"\bhumidity\b", r"\bghi\b", r"\bradiation\b", r"\birradiance\b",
# Sensors
r"\bsensor\b", r"\bsoil\b", r"\bmoisture\b", r"\bleaf temp\b",
r"\bpar\b", r"\bndvi\b", r"\bcwsi\b", r"\bvpd\b",
# Photosynthesis / predictions
r"\bphotosynthesis\b", r"\bassimilation\b", r"\bpredict\b",
r"\bforecast\b", r"\bA rate\b", r"\bcarbon\b",
# Energy
r"\benergy\b", r"\bkwh\b", r"\bpower\b", r"\bgeneration\b",
r"\binverter\b",
# Irrigation
r"\birrigat\b", r"\bwater\b",
# Shading โ€” action-oriented
r"\bshade\b", r"\bshading\b", r"\btilt\b", r"\bangle\b", r"\bpanel\b",
# Temporal / current state
r"\bright now\b", r"\bcurrent\b", r"\btoday\b", r"\btomorrow\b",
r"\byesterday\b", r"\bthis week\b", r"\blast \d+ (hour|day|minute)",
# Direct data ask
r"\bshow me\b", r"\bwhat is\b", r"\bwhat are\b", r"\bhow much\b",
r"\bcheck\b", r"\bstatus\b", r"\bstate\b",
# Hebrew keywords (common farmer queries)
r"ืœื”ืฆืœื™ืœ", r"ื”ืฆืœืœื”", r"ื˜ืžืคืจื˜ื•ืจื”", r"ืžื–ื’ ืื•ื•ื™ืจ", r"ื’ืฉื", r"ืจื•ื—",
r"ืœื—ื•ืช", r"ืงืจื™ื ื”", r"ื”ืฉืงื™ื”", r"ืžื™ื", r"ืื ืจื’ื™ื”", r"ื—ืฉืžืœ",
r"ืขื›ืฉื™ื•", r"ื”ื™ื•ื", r"ืžื—ืจ", r"ืืชืžื•ืœ", r"ืžื” ื”ืžืฆื‘", r"ื›ืžื”",
]
# Compile once
_DATA_PATTERNS = [re.compile(p, re.IGNORECASE) for p in _DATA_KEYWORDS]
# Keywords for pure knowledge / biology rule questions (no tool needed)
_KNOWLEDGE_KEYWORDS = [
r"\bwhy\b.*\brule\b", r"\bexplain\b.*\brule\b",
r"\bwhat is rubisco\b", r"\bwhat is fvcb\b", r"\bwhat is farquhar\b",
r"\btell me about\b.*\bbiology\b", r"\bhow does photosynthesis work\b",
r"\bwhat does .* mean\b",
]
_KNOWLEDGE_PATTERNS = [re.compile(p, re.IGNORECASE) for p in _KNOWLEDGE_KEYWORDS]
@dataclass
class QueryClass:
"""Result of query classification."""
requires_data: bool # True = tool call is mandatory
category: str # "data", "knowledge", "greeting", "ambiguous"
matched_keywords: list[str] = field(default_factory=list)
def classify_query(user_message: str) -> QueryClass:
"""Classify whether a user query requires tool-grounded data."""
msg = user_message.strip()
# Very short / greeting
if len(msg) < 5 or re.match(r"^(hi|hello|hey|thanks|thank you|ok|bye)\b", msg, re.I):
return QueryClass(requires_data=False, category="greeting")
# Check knowledge patterns first (more specific)
for pat in _KNOWLEDGE_PATTERNS:
if pat.search(msg):
return QueryClass(requires_data=False, category="knowledge")
# Check data patterns
matched = []
for pat in _DATA_PATTERNS:
m = pat.search(msg)
if m:
matched.append(m.group())
if matched:
# If the only match is a generic question word ("what is", "show me")
# but no domain-specific data keyword, treat as ambiguous
domain_matches = [m for m in matched if m.lower() not in
{"what is", "what are", "show me", "how much", "check", "status", "state"}]
if not domain_matches:
return QueryClass(requires_data=False, category="ambiguous")
return QueryClass(requires_data=True, category="data", matched_keywords=matched)
# Default: ambiguous โ€” allow LLM to decide
return QueryClass(requires_data=False, category="ambiguous")
# ---------------------------------------------------------------------------
# 2. Response validator โ€” deterministic rule checks
# ---------------------------------------------------------------------------
@dataclass
class RuleViolation:
"""A detected rule violation in a chatbot response."""
rule_name: str
severity: str # "block" or "warn"
message: str
correction: str # What to tell the user instead
def validate_response(
response_text: str,
action: Optional[str] = None,
context: Optional[dict] = None,
) -> list[RuleViolation]:
"""
Check a chatbot response for rule violations.
Parameters
----------
response_text : str
The chatbot's response text.
action : str or None
Extracted action ("shade", "irrigate", "no_action", etc.).
context : dict or None
Current conditions: hour, month, temp_c, stage_id, etc.
Returns
-------
List of RuleViolation objects. Empty list = all good.
"""
violations: list[RuleViolation] = []
ctx = context or {}
text_lower = response_text.lower()
hour = ctx.get("hour")
month = ctx.get("month")
temp_c = ctx.get("temp_c")
stage_id = ctx.get("stage_id")
# Detect if the response recommends shading
_recommends_shade = _text_recommends_shading(text_lower)
# Rule: No shading before NO_SHADE_BEFORE_HOUR
if _recommends_shade and hour is not None and hour < NO_SHADE_BEFORE_HOUR:
violations.append(RuleViolation(
rule_name="no_shade_before_10",
severity="block",
message=f"Response recommends shading before {NO_SHADE_BEFORE_HOUR}:00.",
correction=(
"Morning light is critical for carbon fixation. "
f"Shading should not be recommended before {NO_SHADE_BEFORE_HOUR}:00 regardless "
"of temperature. Panels should remain at full tracking."
),
))
# Rule: No shading in restricted months (unless extreme)
if _recommends_shade and month in NO_SHADE_MONTHS:
# Check if the response mentions extreme conditions
_mentions_extreme = any(w in text_lower for w in [
"extreme", "lethal", "emergency", "severe sunburn", "last resort",
])
if not _mentions_extreme:
violations.append(RuleViolation(
rule_name="no_shade_in_may",
severity="block",
message="Response recommends shading in May without citing extreme conditions.",
correction=(
"May is the flowering/fruit-set period. Shading should be "
"avoided in May unless there is extreme heat causing lethal "
"stress. Panels should remain at full tracking."
),
))
# Rule: Below transition temp shading hurts (RuBP-limited)
if _recommends_shade and temp_c is not None and temp_c < NO_SHADE_TLEAF_BELOW:
violations.append(RuleViolation(
rule_name="temperature_transition",
severity="warn",
message=f"Response recommends shading at {temp_c:.0f}ยฐC (below 28ยฐC transition zone).",
correction=(
f"At {temp_c:.0f}ยฐC, photosynthesis is RuBP-limited โ€” "
f"the vine needs light, not shade. Shading would reduce "
f"photosynthesis. Keep panels at full tracking."
),
))
# Rule: Dormant season โ€” shading is irrelevant, not harmful
if stage_id in ("winter_dormancy",) and _recommends_shade:
violations.append(RuleViolation(
rule_name="no_leaves_no_shade_problem",
severity="warn",
message="Response discusses shading during dormancy.",
correction=(
"The vine is dormant with no leaves. Shading is irrelevant "
"(not harmful, just pointless). Panels should track for "
"maximum energy."
),
))
# Rule: "No shading" answers must explain why
_recommends_no_shade = _text_recommends_no_shading(text_lower)
if _recommends_no_shade:
_has_reason = any(reason in text_lower for reason in [
"light-limited", "rubp", "need light", "needs light",
"full sun", "below 30", "below 28",
"dormant", "no leaves", "no canopy",
"night", "dark", "no radiation", "ghi", "no sun",
"carbon fixation", "morning light",
"not photosynthesi", "not active",
])
if not _has_reason:
violations.append(RuleViolation(
rule_name="no_shading_must_explain",
severity="warn",
message="Response says 'no shading' without explaining why.",
correction=(
"When recommending no shading, always explain the reason: "
"is the vine light-limited (T < 30ยฐC), dormant (no leaves), "
"or is there no radiation? The farmer needs to understand why."
),
))
return violations
# Shared keyword lists for shading detection heuristics
_POSITIVE_SHADE_PHRASES = [
"recommend shading", "should shade", "activate shading",
"tilt the panel", "move the panel", "adjust the panel",
"shade the vine", "shade your vine", "shading would help",
"shading is recommended", "suggest shading", "consider shading",
"apply shading", "deploy shading", "enable shading",
"recommend anti-tracking", "switch to anti-tracking",
]
_NEGATIVE_SHADE_PHRASES = [
"should not shade", "don't shade", "no shading",
"avoid shading", "shading is not", "not recommend shading",
"do not shade", "keep panels tracking", "full tracking",
"shading would reduce", "shading would hurt",
"shading is irrelevant", "shading is unnecessary",
"i would not recommend shading", "i don't recommend shading",
"no shading needed", "shading is not needed",
"no need to shade", "no need for shading",
]
def _text_recommends_shading(text_lower: str) -> bool:
"""Heuristic: does the response recommend activating shade?"""
has_positive = any(p in text_lower for p in _POSITIVE_SHADE_PHRASES)
has_negative = any(p in text_lower for p in _NEGATIVE_SHADE_PHRASES)
# If both present, the negative usually wins (e.g. "some might suggest shading, but I don't recommend it")
return has_positive and not has_negative
def _text_recommends_no_shading(text_lower: str) -> bool:
"""Heuristic: does the response explicitly recommend NOT shading?"""
return any(p in text_lower for p in _NEGATIVE_SHADE_PHRASES)
# ---------------------------------------------------------------------------
# 3. Confidence estimation
# ---------------------------------------------------------------------------
def estimate_confidence(
tool_called: bool,
tool_succeeded: bool,
data_age_minutes: Optional[float],
tool_name: Optional[str] = None,
rule_override: bool = False,
) -> str:
"""
Estimate response confidence based on data grounding.
Returns one of: "high", "medium", "low", "insufficient_data".
"""
# Rule-based override (e.g. dormancy, biology rules) โ€” always high
if rule_override:
return "high"
# No tool called at all
if not tool_called:
return "low" # answering from system prompt / training data only
# Tool was called but failed
if not tool_succeeded:
return "insufficient_data"
# Tool succeeded โ€” check data freshness
if data_age_minutes is None:
# Computed result (FvCB, shading sim) โ€” no age concept
return "high"
if data_age_minutes <= 30:
return "high"
elif data_age_minutes <= 120:
return "medium"
else:
return "low"
# ---------------------------------------------------------------------------
# 4. Source tagging helper
# ---------------------------------------------------------------------------
# Map tool names to human-readable data sources
_TOOL_SOURCES = {
"get_current_weather": "IMS Station 43 (Sde Boker)",
"get_weather_history": "IMS Station 43 (Sde Boker)",
"get_vine_state": "ThingsBoard sensors (on-site)",
"get_sensor_history": "ThingsBoard sensors (on-site)",
"calc_photosynthesis": "Farquhar FvCB model (computed)",
"predict_photosynthesis_ml": "ML ensemble (computed)",
"get_ps_forecast": "FvCB day-ahead forecast (computed)",
"simulate_shading": "Shadow model simulation (computed)",
"compare_tilt_angles": "Shadow model simulation (computed)",
"get_daily_schedule": "Shadow model schedule (computed)",
"get_energy_generation": "IMS + analytical model (estimated)",
"get_energy_history": "IMS + analytical model (estimated)",
"predict_energy": "IMS + analytical model (estimated)",
"run_day_ahead_advisory": "Gemini day-ahead advisor",
"explain_biology_rule": "Built-in biology rules",
"get_photosynthesis_3d": "3D scene (computed)",
}
def get_source_label(tool_name: str) -> str:
"""Return a human-readable source label for a tool."""
return _TOOL_SOURCES.get(tool_name, tool_name)
def tag_tool_result(tool_name: str, tool_result: dict) -> dict:
"""
Add source metadata to a tool result before sending to Gemini.
The tagged result helps Gemini cite sources in its response.
"""
tagged = dict(tool_result)
tagged["_source"] = get_source_label(tool_name)
tagged["_tool"] = tool_name
# Extract data age if present
age = tool_result.get("age_minutes")
if age is not None:
tagged["_data_age_minutes"] = age
if age > 60:
tagged["_freshness_warning"] = (
f"This data is {age:.0f} minutes old. "
"Warn the user that conditions may have changed."
)
# Validate numeric ranges โ€” flag physically impossible values
range_warnings = validate_numeric_ranges(tool_name, tool_result)
if range_warnings:
tagged["_range_warnings"] = range_warnings
return tagged
# ---------------------------------------------------------------------------
# 5. Numeric range validation โ€” catch sensor faults & model errors
# ---------------------------------------------------------------------------
# Physical bounds for common fields (field_name โ†’ (min, max, unit))
_PHYSICAL_BOUNDS: dict[str, tuple[float, float, str]] = {
"air_temperature_c": (-10.0, 55.0, "ยฐC"),
"ghi_w_m2": (0.0, 1400.0, "W/mยฒ"),
"rh_percent": (0.0, 100.0, "%"),
"wind_speed_ms": (0.0, 50.0, "m/s"),
"A_net": (-5.0, 40.0, "ยตmol COโ‚‚/mยฒ/s"),
"power_kw": (0.0, 60.0, "kW"),
"daily_kwh": (0.0, 500.0, "kWh"),
"PAR": (0.0, 2500.0, "ยตmol/mยฒ/s"),
"Tleaf": (-5.0, 60.0, "ยฐC"),
"VPD": (0.0, 10.0, "kPa"),
"CO2": (200.0, 800.0, "ppm"),
"CWSI": (0.0, 1.0, ""),
"staleness_minutes": (0.0, 1440.0, "min"),
}
def validate_numeric_ranges(tool_name: str, result: dict) -> list[str]:
"""Check tool result values against physical bounds.
Returns a list of warning strings for out-of-range values.
"""
warnings: list[str] = []
for key, (lo, hi, unit) in _PHYSICAL_BOUNDS.items():
val = result.get(key)
if val is None:
continue
try:
v = float(val)
except (TypeError, ValueError):
continue
if v < lo or v > hi:
warnings.append(
f"{key}={v:.1f}{unit} is outside physical range "
f"[{lo:.0f}โ€“{hi:.0f}] โ€” possible sensor fault"
)
return warnings
# ---------------------------------------------------------------------------
# 6. Cross-source consistency check
# ---------------------------------------------------------------------------
def check_cross_source_consistency(
weather: Optional[dict],
sensors: Optional[dict],
) -> list[str]:
"""Compare IMS weather and TB sensor readings for consistency.
Returns a list of caveat strings when sources diverge significantly.
"""
caveats: list[str] = []
if not weather or not sensors:
return caveats
if "error" in weather or "error" in sensors:
return caveats
# Temperature: IMS air temp vs TB treatment air temp
ims_temp = weather.get("air_temperature_c")
tb_temp = sensors.get("treatment_air_temp_c")
if ims_temp is not None and tb_temp is not None:
try:
diff = abs(float(ims_temp) - float(tb_temp))
if diff > 5.0:
caveats.append(
f"IMS air temperature ({float(ims_temp):.1f}ยฐC) and on-site sensor "
f"({float(tb_temp):.1f}ยฐC) differ by {diff:.1f}ยฐC โ€” one source may "
f"be stale or malfunctioning."
)
except (TypeError, ValueError):
pass
return caveats