Agentic RAG、GraphRAG 与 RuleRAG：构建可追溯业务证据链

1
Canonical Source
2
  原始政策、合同、知识文档、数据库事实
3

4
Retrieval Projection
5
  chunk、embedding、sparse index、graph、summary
6

7
Evidence Package
8
  针对当前任务选出的可引用、可验证候选

1
from datetime import datetime
2
from typing import Literal
3
from pydantic import BaseModel
4

5

6
class SourceManifest(BaseModel):
7
    document_id: str
8
    source_uri: str
9
    source_type: Literal["policy", "manual", "faq", "contract", "ticket"]
10
    tenant_id: str
11
    title: str
12
    version: str
13
    content_hash: str
14
    effective_from: datetime | None
15
    effective_to: datetime | None
16
    jurisdiction: str | None
17
    product_scope: list[str]
18
    acl: list[str]
19
    parser_version: str
20
    ingestion_run_id: str
21
    status: Literal["staged", "published", "superseded", "deleted"]

1
discover source
2
-> parse
3
-> validate
4
-> build all projections
5
-> run quality checks
6
-> create immutable index snapshot
7
-> atomically switch active snapshot

1
页数/段落数是否异常
2
标题层级是否保留
3
表格行列是否错位
4
脚注/例外条款是否丢失
5
OCR 置信度
6
重复页与页眉页脚
7
字符编码
8
来源 hash

1
{
2
  "chunk_id": "refund-v18-R001-C2",
3
  "document_id": "refund-policy-v18",
4
  "section_path": ["退款", "未发货", "例外"],
5
  "parent_chunk_id": "refund-v18-R001",
6
  "previous_chunk_id": "refund-v18-R001-C1",
7
  "next_chunk_id": "refund-v18-R001-C3"
8
}

1
原始 chunk：
2
  超过 48 小时仍未发货时，用户可申请全额退款。
3

4
索引上下文：
5
  来自《平台普通商品退款规则》v18，章节“未发货退款”；
6
  适用于非预售普通商品，跨境商品另见 R018。

1
表格：保留表名、表头、行键和单元格坐标
2
图片：保存原图引用、OCR/视觉描述与置信度
3
代码：按符号/函数/类切分，保留文件与行号
4
公式：保留原始表达和变量定义

1
class EvidenceChunk(BaseModel):
2
    chunk_id: str
3
    document_id: str
4
    document_version: str
5
    text: str
6
    content_hash: str
7
    source_uri: str
8
    section_path: list[str]
9
    tenant_id: str
10
    acl: list[str]
11
    jurisdiction: str | None
12
    product_scope: list[str]
13
    effective_from: datetime | None
14
    effective_to: datetime | None
15
    published_at: datetime
16
    language: str
17
    embedding_model: str
18
    embedding_version: str
19
    index_snapshot: str

1
authenticated identity
2
-> allowed tenant/namespaces
3
-> ACL/security classification
4
-> valid-time/jurisdiction/product
5
-> candidate retrieval
6
-> relevance ranking

1
effective_from <= :decision_time
2
AND (effective_to IS NULL OR effective_to > :decision_time)

1
规则编号 R001
2
订单状态枚举 not_shipped
3
专有名词
4
精确错误码
5
罕见关键词

1
口语与规范文本语义映射
2
同义表达
3
概念相关
4
跨表述召回

1
Sparse Top-N ─┐
2
              ├-> Fusion -> Candidate Pool
3
Dense Top-N ──┘

1
RRF(d) = Σ 1 / (k + rank_i(d))

1
Candidate 100
2
-> cross-encoder / late interaction / LLM reranker
3
-> Evidence 10

1
定义
2
父条款
3
前后条件
4
例外
5
表头

1
三天了还没发货，能退吗？

1
{
2
  "topic": "refund_eligibility",
3
  "facts": {
4
    "order_status": "paid",
5
    "logistics_status": "not_shipped",
6
    "hours_since_payment": 72
7
  },
8
  "required_evidence": [
9
    "eligibility_rule",
10
    "merchant_exception",
11
    "refund_amount_limit"
12
  ],
13
  "decision_time": "2026-07-15T10:00:00Z"
14
}

1
跨境预售订单 72 小时没发货，优惠券会退吗，还能补偿吗？

1
Q1 未发货退款资格
2
Q2 跨境/预售例外
3
Q3 优惠券退回规则
4
Q4 补偿资格与金额上限
5
Q5 多动作互斥/顺序

1
Required Constraint Preservation
2
Entity/ID Preservation
3
Negation Preservation
4
Temporal Constraint Accuracy
5
Subquery Coverage
6
Retrieval Recall Delta

1
analyze_need
2
  -> retrieve
3
  -> grade_candidates
4
     -> sufficient -> synthesize
5
     -> irrelevant -> rewrite
6
     -> incomplete -> decompose
7
     -> missing_fact -> call_tool/ask_user
8
     -> conflict -> retrieve_primary_source/human
9
  -> verify_citations
10
  -> answer/decision

1
class RetrievalState(BaseModel):
2
    retrieval_run_id: str
3
    goal: str
4
    query_plan: list[str]
5
    executed_queries: list[str]
6
    evidence_refs: list[str]
7
    missing_requirements: list[str]
8
    conflict_groups: list[str]
9
    iterations: int
10
    retrieval_calls: int
11
    rerank_calls: int
12
    tokens_used: int
13
    stop_reason: str | None

1
required evidence covered
2
no new unique evidence
3
same query repeated
4
max iterations
5
retrieval/token/cost deadline
6
primary sources conflict
7
permission blocks required evidence
8
human decision required

1
class EvidenceGrade(BaseModel):
2
    relevant: bool
3
    supports_requirements: list[str]
4
    missing_requirements: list[str]
5
    source_authority: str
6
    temporal_validity: str
7
    conflicts_with: list[str]
8
    next_action: Literal[
9
        "sufficient",
10
        "rewrite",
11
        "decompose",
12
        "fetch_primary",
13
        "ask_user",
14
        "stop_insufficient",
15
    ]

1
class RetrievalProgress(BaseModel):
2
    query_fingerprint: str
3
    candidate_set_hash: str
4
    new_evidence_count: int
5
    newly_covered_requirements: list[str]
6
    repeated_query_count: int

1
连续两轮 candidate_set_hash 相同
2
连续两轮没有覆盖新 requirement
3
rewrite 与历史 query 高度重复
4
同一冲突只重复召回双方

1
每个 subquery 独立记录结果与错误
2
按 evidence_id/content_hash 去重
3
同一来源不同版本不合并
4
保留每个候选被哪些 query 命中
5
部分子查询失败时标记 missing requirement

1
source URI
2
document/version
3
section/span
4
exact quote or normalized excerpt
5
content hash
6
index snapshot

1
{
2
  "claim_id": "C1",
3
  "claim": "超过48小时未发货可申请全额退款",
4
  "evidence_ids": ["E-R001-1"],
5
  "support": "entailed",
6
  "citation_span_valid": true
7
}

1
{
2
  "status": "insufficient",
3
  "known_facts": ["订单已支付72小时"],
4
  "missing": ["是否属于预售例外"],
5
  "next_action": "ask_user_or_query_order_product_type"
6
}

1
Nodes
2
  Entity / Document / Section / Event / Rule / Concept
3

4
Edges
5
  mentions / belongs_to / depends_on / exception_to /
6
  supersedes / applies_to / caused_by / supported_by
7

8
Community / Hierarchy
9
  对密集子图生成分层摘要

1
某个商家有哪些关联投诉与规则？
2
规则 R001 有哪些例外和依赖定义？
3
订单事件与物流节点如何关联？

1
近一个季度退款投诉的主要主题是什么？
2
规则库有哪些跨文档冲突模式？

1
语料小且问题只是局部事实
2
实体/关系提取质量低
3
知识更新极快但图维护链路不成熟
4
没有需要多跳或全局汇总的查询
5
团队无法运维图、社区摘要和版本

1
edge_id
2
source_refs
3
extractor_version
4
confidence
5
valid_from/valid_to
6
recorded_at
7
status

1
“上海商家 A”
2
“A 商贸（上海）有限公司”
3
“merchant_id=8921”

1
class EntityMergeProposal(BaseModel):
2
    left_entity_id: str
3
    right_entity_id: str
4
    matching_signals: list[str]
5
    conflicting_signals: list[str]
6
    confidence: float
7
    source_refs: list[str]

1
识别受影响 document/chunk/entity/edge
2
-> tombstone 旧派生边
3
-> 写入新版本边
4
-> 重算受影响社区
5
-> 重建相关层级摘要
6
-> 发布 graph snapshot

1
适用主体
2
前置条件
3
结论/动作
4
例外
5
优先级
6
地域/产品范围
7
生效与失效时间
8
与其他规则关系

1
class RuleRecord(BaseModel):
2
    rule_id: str
3
    version: str
4
    title: str
5
    jurisdiction: str
6
    product_scope: list[str]
7
    effective_from: datetime
8
    effective_to: datetime | None
9
    priority: int
10
    condition_ast: dict
11
    outcome: dict
12
    exception_rule_ids: list[str]
13
    depends_on_rule_ids: list[str]
14
    source_uri: str
15
    source_quote: str
16
    content_hash: str
17
    extraction_status: Literal["draft", "human_verified", "published"]

1
{
2
  "all": [
3
    {"field": "order.status", "op": "eq", "value": "paid"},
4
    {"field": "logistics.status", "op": "eq", "value": "not_shipped"},
5
    {"field": "hours_since_payment", "op": "gte", "value": 48},
6
    {"not": {"field": "product.type", "op": "eq", "value": "presale"}}
7
  ]
8
}

1
Verified Business Facts
2
  -> Build Rule Retrieval Query
3
  -> Hybrid Retrieve Candidate Rules
4
  -> Expand Definitions/Exceptions/Dependencies
5
  -> Filter Published + Effective + Jurisdiction + Product
6
  -> Deterministic Condition Evaluation
7
  -> Resolve Priority/Conflict
8
  -> Evidence Bundle
9
  -> LLM Explain Result
10
  -> Policy Engine Recheck Before Action

1
更具体地域/产品规则
2
> 更高显式 priority
3
> 更新且当前生效版本
4
> 一般规则

1
Natural-language Policy
2
  -> Clause Segmentation
3
  -> LLM Rule Draft
4
  -> Schema Validation
5
  -> Symbol/Field Resolution
6
  -> Example + Counterexample Generation
7
  -> Deterministic Test Execution
8
  -> Human Policy-owner Review
9
  -> Signed Published Rule Snapshot

1
{
2
  "rule_id": "R001",
3
  "status": "blocked",
4
  "unresolved_symbols": ["承诺发货时间"],
5
  "missing_definitions": ["预售商品"],
6
  "required_review": "policy_owner"
7
}

1
{
2
  "rule_id": "R001",
3
  "tests": [
4
    {"hours": 47, "product_type": "normal", "expected": false},
5
    {"hours": 48, "product_type": "normal", "expected": true},
6
    {"hours": 72, "product_type": "presale", "expected": false},
7
    {"hours": 72, "product_type": "normal", "order_status": "cancelled", "expected": false}
8
  ]
9
}

1
R001：普通商品支付超过 48 小时仍未发货，用户可申请全额退款。
2
R001-E1：预售商品按承诺发货时间判断，不适用 48 小时规则。
3
R002：物流连续 7 天无更新，标记物流异常并允许转人工处理。
4
R004：平台自动补偿不得超过实付金额的 10%，且最高 5000 分。

1
{
2
  "order.id": "ORD-A1B2C3D4",
3
  "order.status": "paid",
4
  "order.paid_amount_cents": 12900,
5
  "product.type": "normal",
6
  "logistics.status": "not_shipped",
7
  "hours_since_payment": 72,
8
  "fact_versions": {
9
    "order": 8,
10
    "logistics": 3
11
  }
12
}

1
{
2
  "candidate_rules": ["R001", "R001-E1", "R004"],
3
  "evaluations": [
4
    {
5
      "rule_id": "R001",
6
      "result": "matched",
7
      "satisfied_conditions": [
8
        "order.status == paid",
9
        "logistics.status == not_shipped",
10
        "hours_since_payment >= 48",
11
        "product.type != presale"
12
      ]
13
    },
14
    {
15
      "rule_id": "R001-E1",
16
      "result": "not_applicable",
17
      "failed_condition": "product.type == presale"
18
    }
19
  ]
20
}

1
{
2
  "decision": "refund_eligible",
3
  "responsible_party": "merchant",
4
  "allowed_action": "create_refund",
5
  "max_amount_cents": 12900,
6
  "rule_ids": ["R001"],
7
  "fact_refs": [
8
    "order://ORD-A1B2C3D4?version=8",
9
    "logistics://ORD-A1B2C3D4?version=3"
10
  ],
11
  "evidence_bundle_id": "evb_01J...",
12
  "requires_confirmation": true
13
}

1
重查订单 version
2
重查物流状态
3
校验 Rule Snapshot 仍可执行
4
校验确认参数 hash
5
Policy Engine 授权
6
执行幂等退款
7
查询最终状态

1
文档 ACL
2
chunk ACL
3
source connector credential
4
tenant namespace
5
field-level redaction
6
purpose limitation

1
来源 allowlist/quality gate
2
指令与数据显式分隔
3
内容安全标签
4
只抽取当前任务字段
5
文档文本不能修改系统策略
6
工具调用仍走独立授权
7
高风险证据需权威来源/人工

1
恶意文档入库
2
SEO/关键词堆叠劫持 sparse 排名
3
embedding collision/相似文本淹没
4
图中伪造高连接实体
5
社区摘要污染

1
Parse Success Rate
2
Structure Preservation
3
Table/OCR Accuracy
4
Metadata Completeness
5
ACL Accuracy
6
Index Freshness Lag
7
Duplicate Rate

1
Recall@K
2
Precision@K
3
MRR / nDCG
4
Required Evidence Recall
5
Effective-version Accuracy
6
Unauthorized Retrieval Rate
7
No-answer Detection

1
Rerank Lift
2
Evidence Sufficiency Accuracy
3
Conflict Detection Recall
4
Exception Retrieval Recall
5
Evidence Diversity
6
Quote/Span Accuracy

1
Claim Faithfulness
2
Citation Precision/Recall
3
Citation Entailment
4
Unsupported Claim Rate
5
Rule Match Accuracy
6
Action Accuracy
7
Policy Compliance

1
Retrieval Iterations
2
Calls / Tokens / Cost
3
p50/p95/p99 Latency
4
Cache Hit Rate
5
Index Snapshot Age
6
Failure/Retry Rate

1
{
2
  "case_id": "RAG-RF-001",
3
  "query": "普通商品支付72小时没发货，可以退款吗？",
4
  "decision_time": "2026-07-15T10:00:00Z",
5
  "identity": {
6
    "tenant_id": "shop-a",
7
    "roles": ["customer_support"]
8
  },
9
  "required_rule_ids": ["R001"],
10
  "required_exception_ids": ["R001-E1"],
11
  "expected_decision": "refund_eligible",
12
  "forbidden_document_versions": ["refund-policy-v17"]
13
}

1
同义口语
2
编号精确查询
3
多条件分解
4
否定与例外
5
旧版本/未来版本
6
跨地域/跨产品
7
权限拒绝
8
无答案
9
来源冲突
10
注入/投毒
11
图多跳
12
全局主题

1
移除关键规则 -> 应报告证据不足
2
加入高相似旧规则 -> 不应误选
3
加入跨租户更相关文档 -> 必须过滤
4
把普通商品改为预售 -> 决策应改变
5
把 72 小时改为 24 小时 -> R001 不匹配

1
Offline：固定 Snapshot、可重复、适合回归
2
Shadow：真实流量不影响用户，比较候选链路
3
Canary：小流量真正生效，监控安全与业务指标
4
Online Eval：抽样人工/模型评审 + 环境结果

1
tenant
2
identity/ACL fingerprint
3
normalized query
4
filters
5
index snapshot
6
retriever/reranker version
7
top-k

1
Embedding Cache
2
Retrieval Candidate Cache
3
Evidence Bundle Cache

1
build snapshot N+1
2
-> validate
3
-> publish active pointer
4
-> new runs pin N+1
5
-> old runs finish on N
6
-> retain N for replay window
7
-> retire according to policy

1
mark snapshot revoked
2
Policy Engine blocks high-risk actions from it
3
running agents revalidate before write
4
publish corrected snapshot
5
replay affected decisions

1
{
2
  "retrieval_run_id": "ret_01J...",
3
  "query_plan_version": "qp@7",
4
  "queries": ["普通商品 未发货 48小时 退款 例外"],
5
  "filters": {
6
    "tenant": "shop-a",
7
    "effective_at": "2026-07-15T10:00:00Z",
8
    "product": "normal"
9
  },
10
  "index_snapshot": "policy-2026-07-15.3",
11
  "retrievers": ["bm25@2", "dense@embed-v5"],
12
  "fusion": "rrf@k60",
13
  "reranker": "cross-encoder@4",
14
  "candidate_count": 83,
15
  "evidence_count": 5,
16
  "stop_reason": "requirements_covered"
17
}

1
ingestion parser/OCR cost per document
2
embedding cost per published chunk
3
graph extraction and community summary cost
4
retrieval/rerank latency per query
5
agentic iteration cost per verified answer
6
storage/index cost per active snapshot