Spaces:
Sleeping
Sleeping
File size: 43,287 Bytes
88245f7 1933348 88245f7 bcd961e 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 9918f43 1933348 9918f43 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 34b037e 1933348 88245f7 1933348 88245f7 1933348 88245f7 bcd961e 88245f7 bcd961e 88245f7 bcd961e 88245f7 1933348 88245f7 bcd961e 88245f7 1933348 88245f7 bcd961e 88245f7 bcd961e 88245f7 bcd961e 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 bcd961e 1933348 88245f7 1933348 bcd961e 88245f7 bcd961e 1933348 bcd961e 1933348 bcd961e 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 bcd961e 88245f7 bcd961e 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 9918f43 1933348 88245f7 bcd961e 1933348 bcd961e 1933348 88245f7 bcd961e 88245f7 bcd961e 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 9918f43 88245f7 bcd961e 88245f7 1933348 88245f7 1933348 88245f7 9918f43 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 88245f7 1933348 bcd961e 88245f7 bcd961e 88245f7 1933348 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 | """Agent implementations.
Three agents make up the system:
* :class:`CoachAgent` β orchestrator. Picks one action per turn
(``call_agent`` / ``ask_user`` / ``write_memory`` / ``compose_response``)
and self-reviews the plan against medical flags before composing.
* :class:`MedicalAssessmentAgent` β runs deterministic clinical formulas
*first*, then asks the LLM to interpret the numbers and emit flags /
recommendations / evidence. The agent overwrites whatever the LLM
emitted for ``calculations`` with the deterministic values so the math
is exact by construction.
* :class:`PlannerAgent` β drafts a meal plan, looks up nutrition via the
WebSearchTool, runs the PuLP solver, and runs :func:`check_plan` on the
result before returning. Up to two internal revisions resolve any
deterministic violations (allergies / calorie or macro deviations)
without involving the Coach.
The public :func:`check_plan` mirrors the Planner's internal critic so the
eval harness can exercise the same code path against hand-built fixtures.
"""
from __future__ import annotations
import json
from datetime import datetime
from typing import Any, Dict, List, Optional, Tuple
from config import get_settings
from logging_setup import get_logger
from nutrition_formulas import full_assessment
from schemas import (
Calculations,
CoachDecision,
MacroTargets,
MedicalAssessmentDecision,
MedicalAssessmentResult,
PlannerDecision,
)
from tools import QuantitiesFinder, WebSearchTool
from utils import save_to_json, should_debug, update_memory_partition
_coach_logger = get_logger("agents.coach")
_medical_logger = get_logger("agents.medical")
_planner_logger = get_logger("agents.planner")
# Tolerances are module-level so tests / configs can override.
CALORIE_TOLERANCE = 0.03 # +/- 3 %
MACRO_TOLERANCE = 0.05 # +/- 5 %
def _decode_plan_field(value: Any) -> Optional[Dict[str, Any]]:
"""Normalise a Planner plan field to a dict.
Structured decoding usually returns a dict, but mocked tests and some
SDK paths pass the same payload as a JSON-encoded string. Both shapes
flow through here, so the agent only ever sees ``Optional[dict]``.
"""
if value is None:
return None
if isinstance(value, dict):
return value
if isinstance(value, str):
text = value.strip()
if not text:
return None
try:
decoded = json.loads(text)
except json.JSONDecodeError:
_planner_logger.warning(
"Planner plan field was not valid JSON; using raw string. Preview: %s",
text[:200],
)
return {"raw": text}
if isinstance(decoded, dict):
return decoded
return {"value": decoded}
return None
def _is_plan_empty(plan: Dict[str, Any]) -> bool:
"""True when ``plan`` contains no renderable content.
Catches the specific anti-pattern ``{"days": [{}]}`` that constrained
decoding can emit when the model fills the envelope but skips the
body. Anything with non-empty entries under ``days`` is accepted β
clinical quality is the post-LP check's job, not this guard's.
"""
if not isinstance(plan, dict):
return True
days = plan.get("days")
if not days:
return True
if isinstance(days, list):
if all(not isinstance(d, (dict, list)) or not d for d in days):
return True
return False
# ---------------------------------------------------------------------------
# Plan checker (public β used by Planner internally and by the eval harness)
# ---------------------------------------------------------------------------
def _extract_foods_and_totals(plan: Dict[str, Any]) -> Tuple[List[Dict[str, Any]], Dict[str, float]]:
"""Walk a plan and collect (list-of-foods, summed-totals).
Supports both the canonical ``{"days": [[FoodItem, ...], ...]}`` shape
and the flatter ``{"days": [{food}, ...]}`` shape the LLM tends to emit
when constrained decoding inlines the inner list. ``daily_totals`` (if
present) overrides the walked sums.
"""
foods: List[Dict[str, Any]] = []
totals: Dict[str, float] = {
"calories": 0.0,
"protein": 0.0,
"fat": 0.0,
"carbohydrates": 0.0,
}
def _walk(node: Any) -> None:
if isinstance(node, list):
for item in node:
_walk(item)
elif isinstance(node, dict):
if "name" in node and any(k in node for k in ("calories", "calories_g", "kcal")):
foods.append(node)
totals["calories"] += float(node.get("calories", node.get("kcal", 0)) or 0)
totals["protein"] += float(node.get("protein_g", node.get("protein", 0)) or 0)
totals["fat"] += float(node.get("fat_g", node.get("fat", 0)) or 0)
totals["carbohydrates"] += float(
node.get("carbohydrates_g", node.get("carbohydrates", 0)) or 0
)
else:
for v in node.values():
_walk(v)
_walk(plan)
if isinstance(plan, dict) and "daily_totals" in plan:
dt = plan["daily_totals"] or {}
for k in ("calories", "protein", "fat", "carbohydrates"):
if k in dt:
totals[k] = float(dt[k])
# Also accept the ``_g`` variants surfaced by the Planner schema.
g_key = f"{k}_g" if k != "calories" else None
if g_key and g_key in dt:
totals[k] = float(dt[g_key])
return foods, totals
def check_plan(plan: Dict[str, Any], memory: Dict[str, Any]) -> List[Dict[str, Any]]:
"""Deterministic safety / target check over a finalised plan.
Returns a list of issue dicts ``{code, description, severity}``. The
Planner uses the severity to drive its internal revision loop:
* ``high`` β allergy violations. Hard block; force a revision.
* ``medium`` β calorie deviation > 3 % or any macro deviation > 5 %.
* ``low`` β disliked-food advisory.
Empty list means the plan passes every deterministic check. The
Coach's self-review covers the LLM-graded checks (medical-flag
respect, citation presence, cultural fit).
"""
issues: List[Dict[str, Any]] = []
user_profile = memory.get("user_profile", {}) or {}
allergies = {
str(a).strip().lower()
for a in user_profile.get("allergies", []) or []
if a
}
dislikes_raw = user_profile.get("food_dislikes", "") or ""
dislikes = {d.strip().lower() for d in str(dislikes_raw).split(",") if d.strip()}
flags = memory.get("flags_and_assessments", {}) or {}
calc = flags.get("calculations", {}) or {}
target_calories = calc.get("daily_target_calories")
macro_targets = calc.get("macro_targets") or {}
foods, totals = _extract_foods_and_totals(plan)
for food in foods:
name = str(food.get("name") or "").lower()
for allergen in allergies:
if allergen and allergen in name:
issues.append(
{
"code": "allergy_violation",
"description": f"Food '{name}' matches allergen '{allergen}'.",
"severity": "high",
}
)
for food in foods:
name = str(food.get("name") or "").lower()
for d in dislikes:
if d and d in name:
issues.append(
{
"code": "disliked_food",
"description": f"Food '{name}' matches user dislike '{d}'.",
"severity": "low",
}
)
if target_calories and totals.get("calories"):
dev = abs(totals["calories"] - target_calories) / target_calories
if dev > CALORIE_TOLERANCE:
issues.append(
{
"code": "calorie_deviation",
"description": (
f"Plan total {totals['calories']:.0f} kcal vs target "
f"{target_calories} kcal ({dev*100:.1f}% deviation)."
),
"severity": "medium",
}
)
macro_map = {"protein_g": "protein", "fat_g": "fat", "carbohydrates_g": "carbohydrates"}
for tgt_key, plan_key in macro_map.items():
target = macro_targets.get(tgt_key)
actual = totals.get(plan_key)
if target and actual:
dev = abs(actual - target) / target
if dev > MACRO_TOLERANCE:
issues.append(
{
"code": f"{plan_key}_deviation",
"description": (
f"{plan_key} total {actual:.0f}g vs target {target}g "
f"({dev*100:.1f}% deviation)."
),
"severity": "medium",
}
)
return issues
# ---------------------------------------------------------------------------
# Coach
# ---------------------------------------------------------------------------
_COACH_SYSTEM_PROMPT = """\
You are the Coach Agent (central orchestrator) of a nutrition Multi-Agent System.
Primary responsibilities:
- Translate user intent into a concrete workflow of response_steps.
- Enforce system rules (MedicalAssessment must complete before Planner runs).
- Decide and perform exactly one action per turn: call_agent, ask_user,
write_memory, or compose_response. There is no call_tool β tools are
invoked exclusively by the worker agents.
Inputs each turn:
- observation (string built from user query + memory + history)
- memory partitions: user_profile, medical_history, flags_and_assessments, plans
- response_steps (list, may be empty on the first turn)
Behaviour rules (mandatory):
1. If response_steps is empty, generate ordered steps (max 6). Each step
must include id, actor, prerequisites, and status "pending".
Standard personal-workflow:
1) Validate required user data (height, weight, age, sex, activity_level,
allergies, goal). If missing -> ask_user.
2) Update memory if the user provided new data [action: write_memory].
3) Call MedicalAssessmentAgent with a task to assess the user.
4) Call PlannerAgent with the relevant task.
5) Self-review the resulting plan; revise via PlannerAgent if needed.
6) compose_response.
2. When calling any agent, set the called step status to "in_progress" and
include prerequisites satisfied by your observation.
3. Only call PlannerAgent if memory.flags_and_assessments has
"assessment_status" of "assessment_complete". If missing, call
MedicalAssessmentAgent first.
4. Self-review (after PlannerAgent returns with memory.plans.current_plan
populated). The deterministic checks β allergy / calorie / macro
tolerances β have already been enforced inside the Planner, so do NOT
re-check the math. Focus on:
a) Medical-flag respect: for each entry in flags_and_assessments.flags
(e.g. "diabetes_risk", "high_ldl"), confirm the plan does not
contain foods that contraindicate the flag.
b) Recommendation coverage: every item in
flags_and_assessments.recommendations should be reflected in the
plan or notes; mention any unaddressed one.
c) Citation presence: clinical claims in the assessment require sources.
If memory.plans.current_plan.sources is empty AND the assessment
had medical flags, that is a revision-worthy gap.
d) Cultural / preference fit: if user_profile.country is set, prefer
regional foods. Major mismatches are revision-worthy; small ones
can be noted in compose_response.
If any (a)-(c) issue is found, call PlannerAgent with task =
"Revise the plan to address: " + each issue joined by "; ".
Cap revisions at 2; on the third attempt, compose_response with the
best plan available and append the unresolved issues as warnings.
5. When new personal data appears in user input, add steps to: propose
memory update (write_memory), call MedicalAssessmentAgent if needed,
re-plan if needed.
6. For any write_memory action, provide the full partition contents in
params.data (not diffs). Top-level partitions are merged; dotted paths
like "plans.current_plan" replace the leaf wholesale.
7. HITL escalation: when
flags_and_assessments.requires_professional_consultation is true,
compose_response MUST append the marker
"<<HITL:CLINICIAN_REVIEW_REQUIRED>>" on its own line at the end and
strongly recommend consulting a clinician before following the plan.
Output JSON shape (enforced by schema):
{
"observation": "...",
"thought": "...",
"response_steps": [ ... ],
"action": "call_agent | ask_user | write_memory | compose_response",
"params": { ... }
}
Required params per action:
- call_agent: {"agent_name": "...", "task": "..."}
- ask_user: {"prompt": "..."}
- write_memory: {"partition": "...", "data": {...}}
- compose_response: {"text": "...markdown..."}
Composition rules:
- When composing the response, extract relevant information from memory
state (calorie target, plan details, dietary restrictions, citations)
in markdown.
- Always include a "Trace" line summarising which agents/tools contributed.
- For high-risk profiles (requires_professional_consultation == true),
end with the HITL marker on its own line.
"""
class CoachAgent:
def __init__(self, llm_instance):
self.llm = llm_instance
def handle_task(self, state: Dict[str, Any]) -> Dict[str, Any]:
settings = get_settings()
memory_str = json.dumps(state["memory"], indent=2, default=str)
response_steps = state.get("response_steps", [])
response_steps_str = (
json.dumps(response_steps, indent=2, default=str) if response_steps else "None"
)
truncated_history: List[Dict[str, str]] = []
for msg in state["conversation_history"]:
if msg["role"] == "assistant" and len(msg["content"]) > 200:
truncated_history.append(
{
"role": "assistant",
"content": msg["content"][:200] + "... (full response in memory)",
}
)
else:
truncated_history.append(msg)
history_str = "\n".join(f"{m['role']}: {m['content']}" for m in truncated_history)
observation = (
f"User query: {state['user_question']}\n"
f"Memory State: {memory_str}\n"
f"Current Response Steps: {response_steps_str}\n"
f"Previous Tool Result: {state.get('agent_result', 'None')}\n"
f"Conversation history: {history_str}"
)
prompt = f"{_COACH_SYSTEM_PROMPT}\n\n--- Current State ---\n{observation}"
if should_debug("agents", "CoachAgent"):
_coach_logger.debug("--- Coach Agent Turn %d ---", state["num_turns"] + 1)
if settings.debug_level == "full":
_coach_logger.debug("Raw LLM input:\n%s", prompt)
decision = self.llm.call_typed(prompt, CoachDecision)
if decision is None:
return self._fallback_state(state, "Coach decision could not be parsed.")
if should_debug("agents", "CoachAgent"):
_coach_logger.debug("Coach decision:\n%s", decision.model_dump_json(indent=2))
if not settings.debug_mode:
self._log_user_mode_action(decision)
current_action = {"action": decision.action, "params": decision.params}
new_steps = [s.model_dump() for s in decision.response_steps] or state.get(
"response_steps", []
)
save_to_json(
{
"prompt": prompt,
"decision": decision.model_dump(),
"timestamp": datetime.now().isoformat(),
},
f"coach_agent_{datetime.now().isoformat()}.json",
subdirectory="CoachAgent",
)
return {
**state,
"current_action": current_action,
"response_steps": new_steps,
"num_turns": state["num_turns"] + 1,
"agent_result": None,
}
@staticmethod
def _log_user_mode_action(decision: CoachDecision) -> None:
params = decision.params or {}
action = decision.action
if action == "call_agent":
msg = f"Calling {params.get('agent_name')} with task '{params.get('task')}'"
elif action == "ask_user":
msg = f"Asking user: {params.get('prompt')}"
elif action == "write_memory":
msg = f"Writing to memory partition '{params.get('partition')}'"
elif action == "compose_response":
msg = "Composing final response"
else:
msg = f"Unknown action: {action}"
_coach_logger.info("\nποΈββοΈCoach Agent: %s", msg)
@staticmethod
def _fallback_state(state: Dict[str, Any], message: str) -> Dict[str, Any]:
_coach_logger.error(message)
return {
**state,
"current_action": {
"action": "compose_response",
"params": {
"text": f"Sorry β I hit an internal error while planning. ({message})"
},
"_parse_error": True,
},
"num_turns": state["num_turns"] + 1,
"agent_result": None,
}
# ---------------------------------------------------------------------------
# Medical Assessment
# ---------------------------------------------------------------------------
_MEDICAL_SYSTEM_PROMPT = """\
You are the Medical Assessment Agent. You receive a user's profile +
medical history AND a deterministic computation of BMI / BMR / TDEE /
daily_target_calories / macro_targets that has already been run for you.
Your job is to interpret those numbers and emit clinical flags,
recommendations, and evidence sources β NOT to recompute the math.
Available tools: WebSearchTool (optional, for fetching authoritative
clinical guidelines from WHO / USDA / ADA / EFSA / NICE).
Mandatory behaviour:
1. The system has already computed and given you the calculations block.
Echo it back in result.calculations exactly β do not alter the numbers.
2. Read the user's flags / conditions / medications / lab results and emit:
- flags_to_set: short stable codes (e.g. "diabetes_risk", "high_ldl",
"hypertension_risk"). Use stable, lowercase, snake_case strings.
- recommendations: 3-6 short bullets the Planner can act on.
- requires_professional_consultation: True for medically sensitive
cases (HbA1c > 6.5 %, eGFR concerns, severe hypertension, pregnancy,
active eating disorder, etc.).
3. Optionally call WebSearchTool ONCE to fetch authoritative guidelines
when the user has a flagged condition. Use tool_task = "..." with a
focused question. Capture the URLs in evidence_sources.
4. When complete, set action_type="assessment_complete" and populate
result with assessment_summary + flags_to_set + recommendations +
requires_professional_consultation + calculations + evidence_sources +
trace + data_confidence.
5. If a critical field is missing (age, sex, height, weight,
activity_level), set action_type="ask_user" and list the missing names
in ``fields``.
Output JSON shape (enforced by schema):
{
"medical_reasoning": "...",
"observation": "...",
"risk_assessment_priorities": [...],
"assessment_plan": [...],
"action_type": "call_tool" | "ask_user" | "assessment_complete",
"tool_task": "..." | null, // only when calling WebSearchTool
"fields": [...], // only when ask_user
"result": { ... } // only when assessment_complete
}
"""
# Critical anthropometric fields needed before any computation.
_CRITICAL_FIELDS = ("age", "sex", "height", "weight", "activity_level", "goal")
class MedicalAssessmentAgent:
# 4 iterations is plenty: one LLM call to enrich the assessment, plus
# an optional WebSearchTool round-trip + recovery iterations.
MAX_ITERATIONS = 4
def __init__(self, llm_instance, web_search_tool: WebSearchTool):
self.llm = llm_instance
self.web_search_tool = web_search_tool
def handle_task(self, task: str, memory: Dict[str, Any]) -> str:
_medical_logger.info("\nπ¨π»ββοΈ MEDICAL ASSESSMENT AGENT STARTED")
settings = get_settings()
user_profile = memory.get("user_profile", {}) or {}
# 1. Precheck β short-circuit before paying for an LLM call if we
# don't even have the data to compute on.
missing = [f for f in _CRITICAL_FIELDS if not user_profile.get(f)]
if missing:
msg = (
f"Missing critical fields: {', '.join(missing)}. "
"Please provide the following information to continue the assessment."
)
_medical_logger.info(
"π¨π»ββοΈ MEDICAL ASSESSMENT AGENT: User query needed - %s", msg
)
return msg
# 2. Deterministic clinical math (no LLM, no subprocess).
deterministic_calcs = full_assessment(
weight_kg=float(user_profile["weight"]),
height_cm=float(user_profile["height"]),
age_years=float(user_profile["age"]),
sex=str(user_profile["sex"]),
activity_level=str(user_profile["activity_level"]),
goal=str(user_profile["goal"]),
)
relevant_memory = {
"user_profile": user_profile,
"medical_history": memory.get("medical_history", {}),
}
memory_str = json.dumps(relevant_memory, indent=2, default=str)
calc_str = json.dumps(deterministic_calcs, indent=2, default=str)
tool_results: List[str] = []
assessment_plan: List[dict] = []
# 3. LLM enrichment loop. Usually one iteration; up to MAX in case
# the model wants a WebSearchTool round-trip before finalising.
for iteration in range(self.MAX_ITERATIONS):
tool_results_str = (
"\n".join(f"Tool Result {i+1}: {r}" for i, r in enumerate(tool_results))
or "None"
)
assessment_plan_str = (
json.dumps(assessment_plan, indent=2, default=str)
if assessment_plan
else "None"
)
prompt = (
f"{_MEDICAL_SYSTEM_PROMPT}\n\n--- Task & State ---\n"
f"Task: {task}\n"
f"Current Memory: {memory_str}\n"
f"Deterministic calculations (use these EXACTLY in result.calculations): "
f"{calc_str}\n"
f"Current Assessment Plan: {assessment_plan_str}\n"
f"Previous Tool Results: {tool_results_str}\n"
)
if should_debug("agents", "MedicalAssessmentAgent"):
_medical_logger.debug(
"--- Medical Assessment Iteration %d ---", iteration + 1
)
if settings.debug_level == "full":
_medical_logger.debug("Raw LLM input:\n%s", prompt)
decision = self.llm.call_typed(prompt, MedicalAssessmentDecision)
if decision is None:
_medical_logger.error(
"Medical decision parse failed at iteration %d", iteration + 1
)
# Fall through to the deterministic-only fallback below.
break
if should_debug("agents", "MedicalAssessmentAgent"):
_medical_logger.debug(
"Medical decision:\n%s", decision.model_dump_json(indent=2)
)
if decision.assessment_plan:
assessment_plan = [s.model_dump() for s in decision.assessment_plan]
if not settings.debug_mode:
self._log_user_mode_action(decision)
if decision.action_type == "call_tool":
if not decision.tool_task:
tool_results.append("WebSearchTool: missing tool_task")
continue
tool_results.append(
f"WebSearchTool: {self.web_search_tool.handle_task(decision.tool_task)}"
)
elif decision.action_type == "ask_user":
fields = decision.fields or []
msg = (
f"Missing critical fields: {', '.join(fields)}. "
"Please provide the following information to continue the assessment."
)
_medical_logger.info(
"π¨π»ββοΈ MEDICAL ASSESSMENT AGENT: User query needed - %s", msg
)
return msg
elif decision.action_type == "assessment_complete":
return self._finalize(
task,
decision,
memory,
relevant_memory,
tool_results,
deterministic_calcs,
)
else:
_medical_logger.error("Unknown action_type: %s", decision.action_type)
break
# 4. Deterministic-only fallback. Reached when the LLM never emits
# assessment_complete; we still ship a usable assessment so the
# Coach can proceed instead of stalling.
_medical_logger.warning(
"π¨π»ββοΈ MEDICAL ASSESSMENT AGENT: LLM never finalised; using "
"deterministic-only fallback."
)
return self._finalize_deterministic_only(
task, memory, relevant_memory, tool_results, deterministic_calcs
)
# ------------------------------------------------------------------
@staticmethod
def _log_user_mode_action(decision: MedicalAssessmentDecision) -> None:
if decision.action_type == "call_tool":
_medical_logger.info(
"π¨π»ββοΈ Medical Assessment Agent: Using WebSearchTool for '%s'",
decision.tool_task,
)
elif decision.action_type == "ask_user":
_medical_logger.info(
"π¨π»ββοΈ Medical Assessment Agent: Asking user for missing fields: %s",
", ".join(decision.fields or []),
)
elif decision.action_type == "assessment_complete":
_medical_logger.info("π¨π»ββοΈ Medical Assessment Agent: Completing assessment")
def _finalize(
self,
task: str,
decision: MedicalAssessmentDecision,
memory: Dict[str, Any],
relevant_memory: Dict[str, Any],
tool_results: List[str],
deterministic_calcs: Dict[str, Any],
) -> str:
result: Optional[MedicalAssessmentResult] = decision.result
if result is None:
_medical_logger.error("assessment_complete decision missing result payload")
return self._finalize_deterministic_only(
task, memory, relevant_memory, tool_results, deterministic_calcs
)
# Overwrite whatever calculations the LLM emitted with the
# deterministic values. The LLM may have re-derived (or invented)
# the numbers; only the formula output is allowed to drive the
# Planner.
result.calculations = Calculations(
BMI=deterministic_calcs["BMI"],
BMR=deterministic_calcs["BMR"],
TDEE=deterministic_calcs["TDEE"],
daily_target_calories=deterministic_calcs["daily_target_calories"],
macro_targets=MacroTargets(**deterministic_calcs["macro_targets"]),
)
update_memory_partition(
memory,
"flags_and_assessments",
{
"assessment_summary": result.assessment_summary,
"flags": result.flags_to_set,
"recommendations": result.recommendations,
"requires_professional_consultation": result.requires_professional_consultation,
"calculations": result.calculations.model_dump(),
"evidence_sources": result.evidence_sources,
"data_confidence": result.data_confidence,
"trace": result.trace,
"assessment_status": "assessment_complete",
"assessment_timestamp": datetime.now().isoformat(),
},
)
save_to_json(
{
"task": task,
"memory_input": relevant_memory,
"tool_results": tool_results,
"result": result.model_dump(),
"timestamp": datetime.now().isoformat(),
},
f"medical_assessment_{datetime.now().isoformat()}.json",
subdirectory="MedicalAssessment",
)
_medical_logger.info(
"π¨π»ββοΈ MEDICAL ASSESSMENT AGENT COMPLETED: %s", result.assessment_summary
)
return result.assessment_summary
def _finalize_deterministic_only(
self,
task: str,
memory: Dict[str, Any],
relevant_memory: Dict[str, Any],
tool_results: List[str],
deterministic_calcs: Dict[str, Any],
) -> str:
"""Persist a deterministic-only assessment when the LLM bails.
Lower ``data_confidence`` so the Coach can decide whether to
proceed or escalate; flags are empty because we have no LLM
interpretation to attach.
"""
summary = (
f"Deterministic assessment only (LLM enrichment unavailable). "
f"Daily target: {deterministic_calcs['daily_target_calories']} kcal, "
f"macros: {deterministic_calcs['macro_targets']}."
)
update_memory_partition(
memory,
"flags_and_assessments",
{
"assessment_summary": summary,
"flags": [],
"recommendations": [],
"requires_professional_consultation": False,
"calculations": deterministic_calcs,
"evidence_sources": [],
"data_confidence": 0.6,
"trace": "deterministic-only fallback (no LLM enrichment)",
"assessment_status": "assessment_complete",
"assessment_timestamp": datetime.now().isoformat(),
},
)
save_to_json(
{
"task": task,
"memory_input": relevant_memory,
"tool_results": tool_results,
"deterministic_only": True,
"calculations": deterministic_calcs,
"timestamp": datetime.now().isoformat(),
},
f"medical_assessment_{datetime.now().isoformat()}.json",
subdirectory="MedicalAssessment",
)
return summary
# ---------------------------------------------------------------------------
# Planner
# ---------------------------------------------------------------------------
_PLANNER_SYSTEM_PROMPT = """\
You are the Planner Agent. Create personalised meal plans constrained by
the medical assessment.
Available tools: WebSearchTool, QuantitiesFinder.
Mandatory behaviour:
1. Precondition: do NOT plan unless flags_and_assessments has an
"assessment_status" of "assessment_complete". If missing, return
action_type="provide_plan" with final_plan={"error": "..."} explaining
the blocker and suggesting MedicalAssessmentAgent.
2. Batch tool calls: fetch nutrition facts for ALL foods in one
WebSearchTool call rather than one call per item.
3. For each food in the draft, look up per-100g nutrition (calories,
protein, fat, carbohydrates). If WebSearchTool fails for >2 items,
fall back to internal knowledge.
4. Tolerances: calories +/- 3%, each macro +/- 5% of target. These are
enforced after the QuantitiesFinder solves; if a deviation is flagged
you will receive "Revision issues" in the prompt and must adjust the
draft (swap foods, change anchors) before the next solver call.
5. Exclude allergens and disliked foods. Propose alternatives if necessary.
6. Multi-day requests: emit a 1-2 day plan and instruct the user to rotate.
7. QuantitiesFinder format: tool_task MUST be a JSON STRING containing
{"foods": [...], "targets": {...}}. Each food needs name, calories,
protein, fat, carbohydrates (per 100g) and estimated_g. Targets MUST
use the keys calories, protein, fat, carbohydrates (NOT protein_g
etc.) β the solver validates this strictly.
Planning Steps (used when none provided):
1. Draft a realistic plan; assign a realistic estimated_g per food.
2. Batch-gather nutrition facts via WebSearchTool.
3. Call QuantitiesFinder with foods + targets to compute precise grams.
4. Update the draft with the solver's quantities.
5. Provide the final plan via action_type="provide_plan".
Output JSON shape (enforced by schema):
{
"observation": "...",
"thought": "...",
"planning_steps": [...],
"action_type": "call_tool" | "draft_plan" | "provide_plan",
"tool_name": "WebSearchTool" | "QuantitiesFinder" | null,
"tool_task": "..." | null,
"drafted_plan": { ... } | null,
"final_plan": { ... } | null
}
Final-plan shape (when action_type="provide_plan"):
- final_plan.days MUST be a non-empty list of fully populated food
objects (name, meal_group, grams from the solver, calories, protein_g,
fat_g, carbohydrates_g).
- final_plan.daily_totals mirrors the solver's ``achieved`` block.
- final_plan.sources lists the citation URIs.
- final_plan.trace summarises which agents/tools contributed.
- An empty envelope ({"days": [{}], ...}) is rejected; always transcribe
the solver output before returning.
"""
# Bounded internal revision count. After this, the agent returns whatever
# it has plus an ``unresolved_issues`` block so the Coach can decide.
_MAX_INTERNAL_REVISIONS = 2
class PlannerAgent:
MAX_ITERATIONS = 12
def __init__(
self,
llm_instance,
web_search_tool: WebSearchTool,
quantities_finder: QuantitiesFinder,
):
self.llm = llm_instance
self.web_search_tool = web_search_tool
self.quantities_finder = quantities_finder
def handle_task(self, task: str, memory: Dict[str, Any]) -> str:
_planner_logger.info("\nπ PLANNER AGENT STARTED")
settings = get_settings()
relevant_memory = {
"user_profile": memory.get("user_profile", {}),
"flags_and_assessments": memory.get("flags_and_assessments", {}),
}
tool_results: List[str] = []
planning_steps: List[dict] = []
revision_count = 0
last_issues: List[Dict[str, Any]] = []
for iteration in range(self.MAX_ITERATIONS):
memory_str = json.dumps(
{
"user_profile": memory.get("user_profile", {}),
"flags_and_assessments": memory.get("flags_and_assessments", {}),
"plans": memory.get("plans", {}),
},
indent=2,
default=str,
)
tool_results_str = (
"\n".join(f"Tool Result {i+1}: {r}" for i, r in enumerate(tool_results))
or "None"
)
planning_steps_str = (
json.dumps(planning_steps, indent=2, default=str)
if planning_steps
else "None"
)
revision_block = ""
if last_issues:
revision_block = (
"\n--- Revision issues (must address before next provide_plan) ---\n"
+ "\n".join(
f"- [{i['severity']}] {i['code']}: {i['description']}"
for i in last_issues
)
+ f"\nRevision attempt: {revision_count}/{_MAX_INTERNAL_REVISIONS}\n"
)
prompt = (
f"{_PLANNER_SYSTEM_PROMPT}\n\n--- Task & State ---\n"
f"Task: {task}\n"
f"Current Memory: {memory_str}\n"
f"Current Planning Steps: {planning_steps_str}\n"
f"Previous Tool Results: {tool_results_str}"
f"{revision_block}"
)
if should_debug("agents", "PlannerAgent"):
_planner_logger.debug("--- Planner Iteration %d ---", iteration + 1)
if settings.debug_level == "full":
_planner_logger.debug("Raw LLM input:\n%s", prompt)
decision = self.llm.call_typed(prompt, PlannerDecision)
if decision is None:
_planner_logger.error(
"Planner decision parse failed at iteration %d", iteration + 1
)
return "Planner failed: could not parse LLM decision."
if should_debug("agents", "PlannerAgent"):
_planner_logger.debug(
"Planner decision:\n%s", decision.model_dump_json(indent=2)
)
if decision.planning_steps:
planning_steps = [s.model_dump() for s in decision.planning_steps]
if not settings.debug_mode:
self._log_user_mode_action(decision)
if decision.action_type == "call_tool":
tool_results.append(f"{decision.tool_name}: {self._dispatch_tool(decision)}")
elif decision.action_type == "draft_plan":
drafted = _decode_plan_field(decision.drafted_plan)
if drafted is not None:
memory.setdefault("plans", {})["drafted_plan"] = drafted
tool_results.append("Plan drafted and stored in memory")
else:
tool_results.append("Drafted plan not provided")
elif decision.action_type == "provide_plan":
final = _decode_plan_field(decision.final_plan) or memory.get(
"plans", {}
).get("drafted_plan")
# Error escape hatch (e.g. precondition not met). The
# schema hint advertises an optional ``error`` key, so the
# LLM often emits ``"error": ""`` alongside a valid plan;
# only treat it as a real error when the value is truthy.
if isinstance(final, dict) and final.get("error"):
_planner_logger.error("π PLANNER AGENT ERROR: %s", final["error"])
return json.dumps({"error": final["error"]})
if not final:
tool_results.append("Cannot finalize: missing plan")
continue
if _is_plan_empty(final):
tool_results.append(
"final_plan.days had no food entries. Transcribe the "
"QuantitiesFinder quantities into days[] as fully "
"populated food objects and emit provide_plan again."
)
_planner_logger.warning(
"π Planner Agent: empty final_plan rejected; re-iterating."
)
continue
# Post-LP deterministic check.
issues = check_plan(final, memory)
blocking = [i for i in issues if i["severity"] in {"medium", "high"}]
if blocking and revision_count < _MAX_INTERNAL_REVISIONS:
revision_count += 1
last_issues = issues
_planner_logger.info(
"π Planner Agent: revising (attempt %d/%d) β %d blocking issue(s).",
revision_count,
_MAX_INTERNAL_REVISIONS,
len(blocking),
)
continue
# Either no blocking issues, or we've exhausted internal
# revisions and must return what we have.
memory.setdefault("plans", {})
memory["plans"]["current_plan"] = final
memory["plans"]["plan_timestamp"] = datetime.now().isoformat()
memory["plans"]["revision_count"] = revision_count
memory["plans"]["post_lp_issues"] = issues
memory["plans"].pop("drafted_plan", None)
unresolved = [i for i in issues if i["severity"] == "high"]
envelope = {
"plan": final,
"revisions": revision_count,
"unresolved_issues": unresolved,
}
save_to_json(
{
"task": task,
"memory_input": relevant_memory,
"tool_results": tool_results,
"final_response": decision.model_dump(),
"post_lp_issues": issues,
"revision_count": revision_count,
"timestamp": datetime.now().isoformat(),
},
f"planner_agent_{datetime.now().isoformat()}.json",
subdirectory="PlannerAgent",
)
_planner_logger.info("\nπ PLANNER AGENT COMPLETED")
return json.dumps(envelope)
else:
_planner_logger.error("Unknown action_type: %s", decision.action_type)
break
_planner_logger.warning("π PLANNER AGENT Stopped (MAX ITERATIONS)")
return (
f"Planning stopped after {self.MAX_ITERATIONS} iterations "
f"with {len(tool_results)} actions"
)
# ------------------------------------------------------------------
def _dispatch_tool(self, decision: PlannerDecision) -> str:
tool_name = decision.tool_name
tool_task = decision.tool_task
if not tool_name or not tool_task:
return "Missing tool_name or tool_task"
if tool_name == "WebSearchTool":
return self.web_search_tool.handle_task(tool_task)
if tool_name == "QuantitiesFinder":
return self.quantities_finder.handle_task(tool_task)
return f"Unknown tool: {tool_name}"
@staticmethod
def _log_user_mode_action(decision: PlannerDecision) -> None:
if decision.action_type == "call_tool":
_planner_logger.info(
"π Planner Agent: Using %s for '%s'",
decision.tool_name,
decision.tool_task,
)
elif decision.action_type == "draft_plan":
_planner_logger.info("π Planner Agent: Drafting plan")
elif decision.action_type == "provide_plan":
_planner_logger.info("π Planner Agent: Finalizing plan")
__all__ = [
"CoachAgent",
"MedicalAssessmentAgent",
"PlannerAgent",
"check_plan",
"CALORIE_TOLERANCE",
"MACRO_TOLERANCE",
]
|