File size: 7,447 Bytes
0762fba a688aff 0762fba a688aff 0762fba a688aff 0762fba a688aff 0762fba a688aff 0762fba a688aff 0762fba a688aff 0762fba a688aff 0762fba a688aff 0762fba a688aff 0762fba a688aff 0762fba a688aff 0762fba a688aff 0762fba | 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 | import json
import numpy as np
from backend.db.supabase import get_client
from backend.llm import llm_call, get_embedding
async def handle_agent_query(
company_id: str, scenario: str, context: dict = None, with_brain: bool = True
) -> dict:
if not with_brain:
return await _baseline_query(scenario, context)
db = get_client()
if not db:
return _error_response("Database connection failed.")
res = (
db.table("skills_files")
.select("brain_json")
.eq("company_id", company_id)
.order("compiled_at", desc=True)
.limit(1)
.execute()
)
if not res.data:
return _error_response("No compiled brain found. Please compile first.")
skills = res.data[0]["brain_json"].get("skills", [])
if not skills:
return _error_response("Brain is empty — no skills compiled.")
query_text = f"{scenario} {json.dumps(context or {})}"
query_emb = get_embedding(query_text)
cached = True
for s in skills:
if "embedding_vector" not in s:
cached = False
break
if cached:
skill_embs = np.array([s["embedding_vector"] for s in skills])
query_vec = np.array(query_emb)
norms = np.linalg.norm(skill_embs, axis=1) * np.linalg.norm(query_vec)
norms[norms == 0] = 1e-10
scores = np.dot(skill_embs, query_vec) / norms
top_indices = np.argsort(scores)[-5:][::-1]
scored = []
for idx in top_indices:
scored.append(
{
"skill": skills[idx],
"score": round(float(scores[idx]), 4),
"index": int(idx),
}
)
else:
scored = []
for i, skill in enumerate(skills):
skill_text = f"{skill.get('category', '')} {skill.get('rule', '')} {skill.get('rationale', '')}"
skill_emb = get_embedding(skill_text)
score = float(
np.dot(query_emb, skill_emb)
/ (np.linalg.norm(query_emb) * np.linalg.norm(skill_emb) + 1e-10)
)
scored.append({"skill": skill, "score": round(score, 4), "index": i})
scored.sort(key=lambda x: x["score"], reverse=True)
top_results = scored[:5]
retrieval_scores = [s["score"] for s in top_results]
skills_context = ""
for rank, s in enumerate(top_results):
sk = s["skill"]
skills_context += (
f"\n--- Skill #{rank + 1} (retrieval_score: {s['score']}) ---\n"
)
skills_context += f"Category: {sk.get('category', 'Unknown')}\n"
skills_context += f"Rule: {sk.get('rule', '')}\n"
skills_context += f"Rationale: {sk.get('rationale', '')}\n"
evidence = sk.get("evidence", [])
if isinstance(evidence, list):
skills_context += f"Evidence: {json.dumps(evidence[:3])}\n"
skills_context += f"Compiled Confidence: {sk.get('confidence', 'unknown')}\n"
prompt = """You are a logical policy reasoning engine. Your ONLY job is to compare scenario parameters against rule thresholds using pure arithmetic, then output the correct action.
CRITICAL LANGUAGE INTERPRETATION RULES:
- "No refunds after X days" means: refunds ARE allowed if the scenario is BEFORE X days. The word "after" creates a threshold at X. Below X = allowed. Above X = denied.
- "Full refund within X days" means: refunds are allowed ONLY if scenario is WITHIN X days. Below X = allowed. Above X = denied.
- "No refunds for X" (without a threshold) is an absolute ban.
ALWAYS compute: does the scenario value fall on the ALLOWED side or the DENIED side of the threshold?
Follow these exact steps:
STEP 1: Extract numeric thresholds from the matched rule (e.g., "60 days" → 60).
STEP 2: Extract the corresponding parameter from the scenario (e.g., days_since_purchase=45).
STEP 3: COMPARE: Write the comparison explicitly (e.g., "45 < 60, so customer is BEFORE the threshold").
STEP 4: DECIDE based solely on the comparison outcome.
Example A:
Rule: "No refunds after 60 days. If purchase was more than 60 days ago, deny."
Scenario: days_since_purchase=45
STEP 1: threshold = 60 days
STEP 2: scenario = 45 days
STEP 3: 45 < 60, customer is BEFORE the threshold
STEP 4: Action = approve (customer qualifies under 60-day limit)
Example B:
Rule: "Full refund only within 14 days of purchase"
Scenario: days_since_purchase=45
STEP 1: threshold = 14 days
STEP 2: scenario = 45 days
STEP 3: 45 > 14, customer is AFTER the threshold
STEP 4: Action = deny (outside the refund window)
Your recommended_action MUST exactly match what the math says. Do not let the emotional tone of the rule ("absolutely no", "no exceptions") override the arithmetic threshold.
confidence:
- retrieval_score < 0.3 → 0.0-0.2 (unrelated)
- 0.3-0.5 → 0.2-0.5 (weak)
- 0.5-0.7 → 0.5-0.75 (moderate)
- > 0.7 and correct match → 0.75-0.95 (strong)
- gibberish → 0.0
Respond with ONLY this JSON:
{
"recommended_action": "action based on your math comparison",
"rule_applied": "exact rule text from best matching skill",
"evidence": ["evidence items"],
"skill_matched": "skill category",
"confidence": 0.0,
"reasoning": "STEP 1: [threshold] STEP 2: [scenario value] STEP 3: [numeric comparison] STEP 4: [action]"
}"""
user_content = f"--- Scenario ---\n{scenario}\n\n--- Additional Context ---\n{json.dumps(context or {})}\n\n--- Retrieved Skills (ranked by relevance) ---\n{skills_context}"
response_str = await llm_call(prompt, user_content)
result = _parse_json(response_str)
result["retrieval_scores"] = retrieval_scores
result["cached_embedding"] = cached
return result
async def _baseline_query(scenario: str, context: dict = None) -> dict:
prompt = """You are a generic AI assistant. You have NO company-specific knowledge or policies.
Answer based only on general industry standards. Be honest about your lack of specific context.
Respond with ONLY a JSON object:
{
"recommended_action": "your general recommendation",
"rule_applied": "general industry standard you referenced",
"evidence": [],
"skill_matched": "none",
"confidence": 0.3,
"retrieval_scores": [],
"reasoning": "explain your reasoning, noting you lack company-specific context"
}"""
user_content = f"Scenario: {scenario}\nContext: {json.dumps(context or {})}"
response_str = await llm_call(prompt, user_content)
return _parse_json(response_str)
def _parse_json(raw: str) -> dict:
try:
clean = raw.strip()
if clean.startswith("```json"):
clean = clean[7:]
if clean.startswith("```"):
clean = clean[3:]
if clean.endswith("```"):
clean = clean[:-3]
return json.loads(clean.strip())
except Exception as e:
return {
"recommended_action": "Failed to parse LLM response",
"rule_applied": "none",
"evidence": [],
"skill_matched": "none",
"confidence": 0.0,
"retrieval_scores": [],
"reasoning": f"JSON parse error: {e}. Raw: {raw[:500]}",
}
def _error_response(msg: str) -> dict:
return {
"recommended_action": msg,
"rule_applied": "none",
"evidence": [],
"skill_matched": "none",
"confidence": 0.0,
"retrieval_scores": [],
"reasoning": msg,
}
|