AgentSkillsCN

Compliance Rag

合规 Rag

SKILL.md

Compliance RAG Skill

Purpose

Search and retrieve regulatory guidance, exculpatory evidence, and compliance documentation using Qdrant vector database.

When to Use

  • •Finding payment authorizations and contracts (alibi evidence)
  • •Consulting EU AI Act and AML regulations
  • •Searching adverse media for entity reputation
  • •Retrieving payment grids and approved vendor lists

Available Tools

1. search_alibi

Search for legitimate business explanations in embedded documents.

python
from backend.tools.tools_docs import search_alibi

result = search_alibi.invoke({
    "query": "payment authorization for consulting services",
    "entity_name": "Shell Company Alpha"
})

# Returns:
# {
#     "has_alibi": True,
#     "evidence": [
#         {"content": "MSA dated 2024-01-15...", "source": "contracts/msa.pdf", "relevance_score": 0.92}
#     ],
#     "exculpatory_strength": 0.85
# }

2. consult_regulation

Retrieve relevant regulatory guidance for compliance context.

python
from backend.tools.tools_docs import consult_regulation

result = consult_regulation.invoke({
    "query": "transparency requirements for high-risk AI systems",
    "regulation_type": "eu_ai_act"
})

# Returns:
# {
#     "regulation": "EU AI Act",
#     "relevant_articles": ["Article 13", "Article 14"],
#     "guidance": "High-risk AI systems must provide..."
# }

3. search_payment_justification

Find authorized payment grids, contracts, and invoices.

python
from backend.tools.tools_docs import search_payment_justification

result = search_payment_justification.invoke({
    "entity_name": "Muller & Sons KG"
})

# Returns:
# {
#     "has_valid_authorization": True,
#     "justifications": [
#         {"content": "Payment Grid Q1 2026...", "source": "payment_grids/2026.xlsx"}
#     ]
# }

4. search_adverse_media

Check for negative news, sanctions, or reputation issues.

python
from backend.tools.tools_docs import search_adverse_media

result = search_adverse_media.invoke({
    "entity_name": "Offshore Holdings LLC"
})

# Returns:
# {
#     "adverse_media_risk": "high" | "medium" | "low",
#     "findings": [
#         {"content": "Company linked to Panama Papers...", "source": "news/2024.json"}
#     ]
# }

Qdrant Collections

fis_evidence

Primary collection for all document embeddings.

FieldTypeDescription
contenttextDocument chunk content
sourcekeywordSource file path
doc_typekeywordcontracts, regulations, payment_grids, adverse_media
entity_refskeyword[]Referenced entity names
embeddingvector[768]Gemini text embedding

Document Types

  1. •Contracts/MSAs - Master Service Agreements, vendor contracts
  2. •Payment Grids - Approved payment schedules, amounts, frequencies
  3. •Regulations - EU AI Act, AML directives, local regulations
  4. •Annual Reports - Company financials, ownership structures
  5. •Adverse Media - News articles, sanctions announcements

Embedding Model

Uses Google Gemini text-embedding-004 for semantic search:

python
from langchain_google_genai import GoogleGenerativeAIEmbeddings

embeddings = GoogleGenerativeAIEmbeddings(
    model="models/text-embedding-004",
    google_api_key=settings.google_api_key
)

Integration with Skeptic Agent

The Skeptic agent uses these tools to build a defense:

python
from backend.agents.skeptic import skeptic_agent

# Skeptic will automatically:
# 1. Search for payment justifications
# 2. Look for alibi evidence
# 3. Consult regulations
# 4. Check adverse media (to confirm no issues)

Risk Reduction Scoring

Evidence FoundRisk Reduction
Valid payment authorization-0.3
Alibi evidence (MSA/contract)-0.2
Within peer group norms-0.15
No adverse media-0.1
Complete entity profile-0.1

EU AI Act Article 13 Compliance

The Skeptic cites regulations to ensure transparency:

python
# Example regulatory citation in verdict
{
    "eu_ai_act_compliance": {
        "article_13_satisfied": True,
        "transparency_statement": "Generated by FIS Swarm AI system",
        "human_oversight_required": True
    }
}

Error Handling

python
result = search_alibi.invoke({"query": "test", "entity_name": "Unknown"})
if not result.get("has_alibi"):
    # No exculpatory evidence found - not necessarily an error
    pass
if "error" in result:
    # Actual Qdrant error
    print(f"Search failed: {result['error']}")

Prerequisites

  • •Qdrant running on QDRANT_URL (default: http://localhost:6333)
  • •QDRANT_API_KEY for authentication
  • •Documents pre-loaded via backend/loaders/load_vectors.py