admin/PlantAI-rag

Fork 0

Files

T

potaser 5bb98a937a 初始提交

2026-07-10 17:59:59 +08:00

5.0 KiB

Executable File

Raw Permalink Blame History

多智能体分析与改进方案总结

日期: 2026-04-10 20:15 GMT+8

智能体协作完成的工作

智能体1: 分析师 - 问题诊断

发现的核心问题:

问题	严重程度	影响
证据是LLM生成，非原文摘录	🔴 严重	82%证据无效
复合基因名未展开	🔴 严重	覆盖率虚高
Prompt设计缺陷	🟡 中等	准确率仅55%

根因分析:

Prompt要求LLM"生成"证据，而非"摘录"
未强制要求展开复合实体
缺乏严格的证据验证

智能体2: 设计师 - 方案设计

PC的关键设计原则:

证据必须原文摘录

❌ "genes were expressed in roots" (生成)
✅ "SlSut01 was highly expressed in roots" (原文摘录)

复合实体必须展开

❌ SlSut01-SlSut40 --[expressed_in]--> roots
✅ SlSut01 --[expressed_in]--> roots
✅ SlSut02 --[expressed_in]--> roots
... (40个单独关系)

逐句提取，精准定位
- 每句话单独处理
- 避免信息过载
- 证据位置明确

新Prompt设计:

prompt = """Extract gene relationships from the sentence.

CRITICAL RULES:
1. EVIDENCE MUST BE EXACT TEXT from the sentence
2. Use SPECIFIC gene names only, NO ranges like "SlSut01-SlSut40"
3. For "SlSut05, SlSut06, SlSut07, and SlSut08", create 4 SEPARATE relations

FORMAT: gene_id | relation | target | exact_evidence

EXAMPLE:
Sentence: "SlSut05, SlSut06 were on Chr07"
Output:
SlSut05 | located_on | Chr07 | "SlSut05, SlSut06 were on Chr07"
SlSut06 | located_on | Chr07 | "SlSut05, SlSut06 were on Chr07"
"""

智能体3: 编码器 - 实现方案

多智能体提取器架构:

PDF → 分析师(拆分) → 设计师(Prompt) → 编码器(提取) → 验证器(检查) → 总结者(输出)

关键改进:

强制原文摘录

def validate_evidence(relation, original_text):
    if relation.evidence not in original_text:
        return False  # 拒绝
    return True

复合实体展开

def expand_composite_entities(relation):
    if '-' in relation.source:
        # SlSut01-SlSut40 → 40个单独关系
        for gene in range(1, 41):
            yield Relation(f"SlSut{gene:02d}", ...)

三级验证
- 自动验证: 证据存在性
- 附表验证: 染色体位置
- 人工抽样: 准确率检查

智能体4: 验证器 - 验证策略

验证流程:

def validate(relations, original_text):
    # 1. 证据存在性检查
    for r in relations:
        if r.evidence not in original_text:
            reject(r)
    
    # 2. 实体标准化检查
    for r in relations:
        if not is_valid_gene(r.source):
            reject(r)
    
    # 3. 附表对比验证
    for r in relations:
        if r.relation == 'located_on':
            expected = supplementary_table[r.source]
            if r.target != expected:
                flag_for_review(r)
    
    # 4. 人工抽样
    sample = random.sample(relations, 20)
    accuracy = manual_check(sample)
    
    return accuracy

验证标准:

证据在原文中: >95%
实体标准化: 100%
附表一致性: >90%
人工抽样准确率: >90%

智能体5: 总结者 - 最终方案

核心改进总结

1. Prompt改进 ✅

Before:

提取基因关系，提供证据

After:

CRITICAL RULES:
1. EVIDENCE MUST BE EXACT TEXT from the sentence
2. Use SPECIFIC gene names only, NO ranges
3. For multiple genes, create SEPARATE relations

2. 证据验证 ✅

Before: 无验证 After:

if evidence not in original_text:
    reject(relation)

3. 实体展开 ✅

Before: SlSut01-SlSut40 (1个关系) After: 40个单独关系

4. 三级验证 ✅