OpenClaw/context-gatekeeper
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

complete: full ablation + Phase4 quality evaluation + honest blog post

Browse Source 
Phase2 complete ablation (added missing variants):
- Coverage-only: 20% contamination rate (confirms Gate is critical)
- Gate-only: +5.2 tokens vs Full (coverage optimization marginal on clean data)
- -Recency: 0 effect on clean data
- -IDF: 0 effect on clean data

Phase4 end-to-end quality evaluation:
- CGK vs Last-5 across 5 queries:
  * CGK: 42.2 tok, purity=1.000, anchor_recall=0.638, term_cov=0.380, contamination=0
  * Last-5: 67.6 tok, purity=0.280, anchor_recall=0.066, term_cov=0.080, contamination=5
- All quality metrics CGK >> Last-5 on synthetic clean data

Known honest limitations:
- Still no real dialogue data (synthetic 4-topic only)
- No real LLM calls (quality is rule-estimated)
- Parameter sensitivity only on clean data, not noisy real data
This commit is contained in:
Elaina
2026-04-22 22:48:25 +08:00
parent 9e44748f91
commit 97e1ddf138
4 changed files with 826 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

357
experiments/phase2_complete_ablation.py Normal file
View File

@@ -0,0 +1,357 @@
#!/usr/bin/env python3
"""Phase 2 补全: 完整的 Ablation Study（包含缺失的 -recency, -coverage, -gate 变体）"""
import sys, os, json
sys.path.insert(0, os.path.dirname(os.path.dirname(__file__)))
from src.gatekeeper import ContextGatekeeper
def estimate_tokens(text):
if not text: return 0
chinese = sum(1 for c in text if '\u4e00' <= c <= '\u9fff')
english = len([w for w in text.split() if w.isascii()])
return int(chinese * 0.4 + english * 1.3 + len(text) * 0.05)
def measure_prompt_tokens(selected, query):
ctx = "".join(f"用户: {i['user']}\n助手: {i['assistant']}\n\n" for i in selected)
context_tok = estimate_tokens(ctx)
fmt_overhead = int(context_tok * 0.08)
return context_tok + fmt_overhead + estimate_tokens(query)
def evaluate_contamination(selected, target):
text = " ".join(i['user'] + i['assistant'] for i in selected)
found = [t for t in ['Redis', 'asyncio', 'PostgreSQL', 'Git']
if t.lower() in text.lower() and t.lower() != target.lower()]
return len(found) > 0, found
def evaluate_answer_quality(gate, query, target_topic):
"""
端到端答案质量评估：模拟 LLM 在不同上下文下的回答质量
评估指标:
1. 上下文正确性: 选中的块是否都与目标话题相关
2. 上下文完整性: 选中的块是否覆盖了回答所需的关键信息
3. 回答引用正确性: 如果用这些块让 LLM 回答，答案是否会引用错误话题
由于没有真实 LLM 调用，用规则模拟：
- 相关块比例 = 目标话题块数 / 总块数
- 锚点覆盖率 = query 锚点在选中块中的出现率
"""
sel = gate.select(query)
# 统计
total_blocks = len(sel)
topic_blocks = 0
other_topic_texts = []
for item in sel:
text = item['user'] + ' ' + item['assistant']
found_topics = [t for t in ['Redis', 'asyncio', 'PostgreSQL', 'Git']
if t.lower() in text.lower() and t.lower() != target_topic.lower()]
if found_topics:
other_topic_texts.extend(found_topics)
else:
topic_blocks += 1
# query 锚点覆盖率
q_anchors, _ = gate.anchor_extractor.extract_with_deictic(query)
covered_anchors = 0
context_text = ' '.join(i['user'] + i['assistant'] for i in sel)
for a in q_anchors:
if a.lower() in context_text.lower():
covered_anchors += 1
anchor_coverage = covered_anchors / len(q_anchors) if q_anchors else 0
return {
'total_blocks': total_blocks,
'topic_blocks': topic_blocks,
'purity': topic_blocks / total_blocks if total_blocks > 0 else 0,
'other_topics': list(set(other_topic_texts)),
'anchor_coverage': anchor_coverage,
'is_contaminated': len(other_topic_texts) > 0
}
redis_qa = [
("Redis 分布式锁和 RedLock 算法有什么区别？", "RedLock是..."),
("Redis 集群环境下怎么做分布式锁？", "集群下..."),
("Redis 惰性删除和定期删除有什么区别？", "惰性删除..."),
("Redis 的过期 key 对 RDB 快照有什么影响？", "过期key..."),
("Redis 主从复制断线后如何增量同步？", "PSYNC..."),
]
asyncio_qa = [
("asyncio.Task 的 cancel 方法怎么工作的？", "cancel..."),
("asyncio.gather 和 asyncio.wait 的返回结果有什么区别？", "gather..."),
("asyncio 的事件循环怎么启动和停止？", "事件循环..."),
("asyncio.sleep 和 time.sleep 的区别是什么？", "sleep..."),
("asyncio 的 Future 对象怎么获取结果？", "Future..."),
]
pg_qa = [
("PostgreSQL 的 MVCC 机制是怎么保证读不阻塞写的？", "MVCC..."),
("PostgreSQL 的 VACUUM 为什么要定期运行？", "VACUUM..."),
("EXPLAIN ANALYZE 怎么看执行计划？", "EXPLAIN..."),
("PostgreSQL B-tree 索引和 Hash 索引的区别是什么？", "B-tree..."),
("PostgreSQL 的 TOAST 机制是什么？", "TOAST..."),
]
git_qa = [
("Git 的 rebase 和 merge 的区别是什么？", "rebase..."),
("Git reset 的 --soft、--mixed、--hard 有什么区别？", "reset..."),
("Git stash 暂存区和工作目录的区别是什么？", "stash..."),
("Git 的 bisect 怎么用来快速定位 bug？", "bisect..."),
("Git 的 reflog 怎么用来恢复误删的提交？", "reflog..."),
]
TEST_SEQ = [
("问PG", "EXPLAIN ANALYZE 怎么看执行计划？", "PostgreSQL"),
("问Git", "Git 的 rebase 和 merge 有什么区别？", "Git"),
("问Redis", "Redis 惰性删除和定期删除有什么区别？", "Redis"),
("问asyncio", "asyncio.Task 的 cancel 方法怎么工作的？", "asyncio"),
("再问Git", "Git 的 reset 和 revert 的应用场景有什么区别？", "Git"),
]
def build_gate():
g = ContextGatekeeper(token_budget=4000)
for i in range(5):
g.add_turn(redis_qa[i][0], redis_qa[i][1])
g.add_turn(asyncio_qa[i][0], asyncio_qa[i][1])
g.add_turn(pg_qa[i][0], pg_qa[i][1])
g.add_turn(git_qa[i][0], git_qa[i][1])
return g
def run_gate_only(gate, query):
"""Gate-only: 只做话题过滤，不做覆盖优化（直接返回所有召回块）"""
q_anchors, has_deictic = gate.anchor_extractor.extract_with_deictic(query)
switched = gate.topic_gate.is_topic_switch(query, gate._active_topic)
idf_cache = gate.anchor_extractor._idf_cache
if switched:
candidates = gate.blocks[-15:]
else:
candidates = gate.blocks
retrieved = gate.retriever.retrieve(
candidates, q_anchors, top_m=20, idf_cache=idf_cache,
active_topic_anchors=gate._active_topic[0] if gate._active_topic else None,
topic_switched=switched, query_text=query
)
return [{"user": b.user_text, "assistant": b.assistant_text, "turn_id": b.turn_id} for b, _ in retrieved]
def run_coverage_only(gate, query):
"""Coverage-only: 不做话题过滤，直接对所有块做覆盖优化"""
q_anchors, _ = gate.anchor_extractor.extract_with_deictic(query)
# Bypass gate: 强制 switched=False，全部块作为候选
orig_switched = gate.topic_gate.is_topic_switch
def fake_switch(*args, **kwargs):
return False # 强制不做话题过滤
gate.topic_gate.is_topic_switch = fake_switch
# Bypass content-word filter by setting topic_switched=False in retrieve
orig_retrieve = gate.retriever.retrieve
def no_filter_retrieve(blocks, qa, top_m=20, **kwargs):
# 把所有块都放进来，不做内容词过滤
scored = []
idf_cache = kwargs.get('idf_cache', {})
for block in blocks:
s = gate.retriever.score(block, qa, 0.0, idf_cache)
scored.append((block, s))
scored.sort(key=lambda x: x[1], reverse=True)
return scored[:top_m]
gate.retriever.retrieve = no_filter_retrieve
sel = gate.select(query)
gate.topic_gate.is_topic_switch = orig_switched
gate.retriever.retrieve = orig_retrieve
return sel
def run_no_recency(gate, query):
"""-Recency: 移除 recency 权重"""
orig_WEIGHT_RECENT = gate.retriever.WEIGHT_RECENT
gate.retriever.WEIGHT_RECENT = 0.0
sel = gate.select(query)
gate.retriever.WEIGHT_RECENT = orig_WEIGHT_RECENT
return sel
def run_no_idf(gate, query):
"""-IDF: 移除 IDF 加权（所有词 IDF=1.0）"""
orig_idf_cache = gate.anchor_extractor._idf_cache.copy() if hasattr(gate.anchor_extractor, '_idf_cache') else {}
def fake_idf(anchor):
return 1.0 # 固定 IDF=1.0，取消区分度
orig_idf_fn = gate.anchor_extractor.idf
gate.anchor_extractor.idf = fake_idf
# 清空 idf_cache 让所有词都用默认值
gate.anchor_extractor._idf_cache.clear()
sel = gate.select(query)
gate.anchor_extractor.idf = orig_idf_fn
gate.anchor_extractor._idf_cache = orig_idf_cache
return sel
def main():
results = {
'Full CGK': [],
'Gate-only': [], # 无覆盖优化（ChatGPT指出的缺失）
'Coverage-only': [], # 无门控过滤（ChatGPT指出的缺失）
'-Recency': [], # 无近期偏好
'-IDF': [], # 无IDF加权
'-Deictic': [],
'-Exact Match': [],
'-Trim': [],
'Last-5 (baseline)': [],
}
print("="*70)
print("Phase 2 COMPLETE Ablation Study (补全)")
print("="*70)
for label, query, target in TEST_SEQ:
print(f"\n[{label}] {query[:45]}...")
# Full CGK
g = build_gate()
sel = g.select(query)
pt = measure_prompt_tokens(sel, query)
cont, _ = evaluate_contamination(sel, target)
aq = evaluate_answer_quality(g, query, target)
results['Full CGK'].append({'pt': pt, 'cont': cont, 'aq': aq})
print(f" Full CGK: {pt:5.0f} tok, 污染={cont}, 纯度={aq['purity']:.2f}, 锚点覆盖={aq['anchor_coverage']:.2f}")
# Gate-only
g2 = build_gate()
sel2 = run_gate_only(g2, query)
pt2 = measure_prompt_tokens(sel2, query)
cont2, _ = evaluate_contamination(sel2, target)
results['Gate-only'].append({'pt': pt2, 'cont': cont2})
print(f" Gate-only: {pt2:5.0f} tok, 污染={cont2}")
# Coverage-only
g3 = build_gate()
sel3 = run_coverage_only(g3, query)
pt3 = measure_prompt_tokens(sel3, query)
cont3, _ = evaluate_contamination(sel3, target)
results['Coverage-only'].append({'pt': pt3, 'cont': cont3})
print(f" Coverage-only:{pt3:5.0f} tok, 污染={cont3}")
# -Recency
g4 = build_gate()
sel4 = run_no_recency(g4, query)
pt4 = measure_prompt_tokens(sel4, query)
cont4, _ = evaluate_contamination(sel4, target)
results['-Recency'].append({'pt': pt4, 'cont': cont4})
print(f" -Recency: {pt4:5.0f} tok, 污染={cont4}")
# -IDF
g5 = build_gate()
sel5 = run_no_idf(g5, query)
pt5 = measure_prompt_tokens(sel5, query)
cont5, _ = evaluate_contamination(sel5, target)
results['-IDF'].append({'pt': pt5, 'cont': cont5})
print(f" -IDF: {pt5:5.0f} tok, 污染={cont5}")
# -Deictic
g6 = build_gate()
orig = g6.anchor_extractor.extract_with_deictic
def no_d(text):
a, _ = orig(text)
return a, False
g6.anchor_extractor.extract_with_deictic = no_d
sel6 = g6.select(query)
pt6 = measure_prompt_tokens(sel6, query)
cont6, _ = evaluate_contamination(sel6, target)
results['-Deictic'].append({'pt': pt6, 'cont': cont6})
print(f" -Deictic: {pt6:5.0f} tok, 污染={cont6}")
# -Exact Match
g7 = build_gate()
orig_exact = g7.retriever._exact_match
g7.retriever._exact_match = lambda b, qa: 0.0
sel7 = g7.select(query)
pt7 = measure_prompt_tokens(sel7, query)
cont7, _ = evaluate_contamination(sel7, target)
results['-Exact Match'].append({'pt': pt7, 'cont': cont7})
print(f" -Exact Match: {pt7:5.0f} tok, 污染={cont7}")
# -Trim
g8 = build_gate()
g8._trim_blocks_to_query = lambda blocks, qa: blocks
sel8 = g8.select(query)
pt8 = measure_prompt_tokens(sel8, query)
cont8, _ = evaluate_contamination(sel8, target)
results['-Trim'].append({'pt': pt8, 'cont': cont8})
print(f" -Trim: {pt8:5.0f} tok, 污染={cont8}")
# Last-5 baseline
conv = [{'user': redis_qa[i][0], 'assistant': redis_qa[i][1]} for i in range(5)] + \
[{'user': asyncio_qa[i][0], 'assistant': asyncio_qa[i][1]} for i in range(5)] + \
[{'user': pg_qa[i][0], 'assistant': pg_qa[i][1]} for i in range(5)] + \
[{'user': git_qa[i][0], 'assistant': git_qa[i][1]} for i in range(5)]
sel9 = conv[-5:]
pt9 = measure_prompt_tokens(sel9, query)
cont9, _ = evaluate_contamination(sel9, target)
results['Last-5 (baseline)'].append({'pt': pt9, 'cont': cont9})
print(f" Last-5: {pt9:5.0f} tok, 污染={cont9}")
# Summary
print("\n" + "="*70)
print("Ablation Summary (avg over {} queries)".format(len(TEST_SEQ)))
print("="*70)
full_avg = sum(d['pt'] for d in results['Full CGK']) / len(results['Full CGK'])
full_cont = sum(1 for d in results['Full CGK'] if d['cont']) / len(results['Full CGK']) * 100
print(f"\n{'Method':<20} {'Avg Tokens':>12} {'Cont%':>8} {'ΔTokens':>10} {'Notes':<30}")
print("-"*80)
print(f"{'Full CGK':<20} {full_avg:>12.1f} {full_cont:>8.1f} {'—':>10} {'baseline':<30}")
for name in ['Gate-only', 'Coverage-only', '-Recency', '-IDF', '-Deictic', '-Exact Match', '-Trim', 'Last-5 (baseline)']:
data = results[name]
avg_tok = sum(d['pt'] for d in data) / len(data)
cont_pct = sum(1 for d in data if d['cont']) / len(data) * 100
diff = avg_tok - full_avg
notes = ""
if name == 'Gate-only':
notes = "← 关键模块"
elif name == 'Coverage-only':
notes = "← 无门控"
elif name == '-Recency':
notes = "← recency权重→0"
elif name == '-IDF':
notes = "← IDF=1.0固定"
print(f"{name:<20} {avg_tok:>12.1f} {cont_pct:>8.1f} {diff:>+10.1f} {notes:<30}")
# Key findings
print("\n" + "="*70)
print("Key Findings")
print("="*70)
# Compare Gate-only vs Coverage-only
go_avg = sum(d['pt'] for d in results['Gate-only']) / len(results['Gate-only'])
co_avg = sum(d['pt'] for d in results['Coverage-only']) / len(results['Coverage-only'])
go_cont = sum(1 for d in results['Gate-only'] if d['cont']) / len(results['Gate-only']) * 100
co_cont = sum(1 for d in results['Coverage-only'] if d['cont']) / len(results['Coverage-only']) * 100
print(f"\n1. Gate-only vs Coverage-only (最关键的两个 ablated variants):")
print(f" Gate-only: {go_avg:.1f} tokens, 污染率 {go_cont:.0f}%")
print(f" Coverage-only: {co_avg:.1f} tokens, 污染率 {co_cont:.0f}%")
print(f" 结论: 门控过滤(Gate)对污染率的影响{'远大于' if co_cont > go_cont else '相当'}覆盖优化(Coverage)")
# -Recency effect
rec_avg = sum(d['pt'] for d in results['-Recency']) / len(results['-Recency'])
rec_cont = sum(1 for d in results['-Recency'] if d['cont']) / len(results['-Recency']) * 100
print(f"\n2. -Recency (移除 recency 权重):")
print(f" -Recency: {rec_avg:.1f} tokens, 污染率 {rec_cont:.0f}% (vs Full {full_cont:.0f}%)")
# -IDF effect
idf_avg = sum(d['pt'] for d in results['-IDF']) / len(results['-IDF'])
idf_cont = sum(1 for d in results['-IDF'] if d['cont']) / len(results['-IDF']) * 100
print(f"\n3. -IDF (固定所有词 IDF=1.0):")
print(f" -IDF: {idf_avg:.1f} tokens, 污染率 {idf_cont:.0f}% (vs Full {full_cont:.0f}%)")
print(f" 结论: IDF 加权对 token 消耗的影响 {'明显' if abs(idf_avg - full_avg) > 5 else '较小'}")
out_path = os.path.join(os.path.dirname(__file__), 'phase2_complete_ablation_results.json')
with open(out_path, 'w') as f:
json.dump(results, f, indent=2, ensure_ascii=False)
print(f"\nSaved to: {out_path}")
if __name__ == '__main__':
main()

243
experiments/phase2_complete_ablation_results.json Normal file
View File

@@ -0,0 +1,243 @@
{
"Full CGK": [
{
"pt": 16,
"cont": false,
"aq": {
"total_blocks": 1,
"topic_blocks": 1,
"purity": 1.0,
"other_topics": [],
"anchor_coverage": 1.0,
"is_contaminated": false
}
},
{
"pt": 59,
"cont": false,
"aq": {
"total_blocks": 2,
"topic_blocks": 2,
"purity": 1.0,
"other_topics": [],
"anchor_coverage": 1.0,
"is_contaminated": false
}
},
{
"pt": 19,
"cont": false,
"aq": {
"total_blocks": 1,
"topic_blocks": 1,
"purity": 1.0,
"other_topics": [],
"anchor_coverage": 1.0,
"is_contaminated": false
}
},
{
"pt": 56,
"cont": false,
"aq": {
"total_blocks": 1,
"topic_blocks": 1,
"purity": 1.0,
"other_topics": [],
"anchor_coverage": 1.0,
"is_contaminated": false
}
},
{
"pt": 61,
"cont": false,
"aq": {
"total_blocks": 9,
"topic_blocks": 5,
"purity": 0.5555555555555556,
"other_topics": [
"Redis",
"asyncio"
],
"anchor_coverage": 0.4,
"is_contaminated": true
}
}
],
"Gate-only": [
{
"pt": 16,
"cont": false
},
{
"pt": 59,
"cont": false
},
{
"pt": 45,
"cont": false
},
{
"pt": 56,
"cont": false
},
{
"pt": 61,
"cont": false
}
],
"Coverage-only": [
{
"pt": 16,
"cont": false
},
{
"pt": 32,
"cont": false
},
{
"pt": 19,
"cont": false
},
{
"pt": 18,
"cont": false
},
{
"pt": 128,
"cont": true
}
],
"-Recency": [
{
"pt": 16,
"cont": false
},
{
"pt": 59,
"cont": false
},
{
"pt": 19,
"cont": false
},
{
"pt": 56,
"cont": false
},
{
"pt": 61,
"cont": false
}
],
"-IDF": [
{
"pt": 16,
"cont": false
},
{
"pt": 59,
"cont": false
},
{
"pt": 19,
"cont": false
},
{
"pt": 56,
"cont": false
},
{
"pt": 61,
"cont": false
}
],
"-Deictic": [
{
"pt": 16,
"cont": false
},
{
"pt": 59,
"cont": false
},
{
"pt": 19,
"cont": false
},
{
"pt": 56,
"cont": false
},
{
"pt": 61,
"cont": false
}
],
"-Exact Match": [
{
"pt": 16,
"cont": false
},
{
"pt": 59,
"cont": false
},
{
"pt": 19,
"cont": false
},
{
"pt": 56,
"cont": false
},
{
"pt": 61,
"cont": false
}
],
"-Trim": [
{
"pt": 16,
"cont": false
},
{
"pt": 59,
"cont": false
},
{
"pt": 19,
"cont": false
},
{
"pt": 56,
"cont": false
},
{
"pt": 61,
"cont": false
}
],
"Last-5 (baseline)": [
{
"pt": 70,
"cont": true
},
{
"pt": 72,
"cont": false
},
{
"pt": 71,
"cont": true
},
{
"pt": 71,
"cont": true
},
{
"pt": 74,
"cont": false
}
]
}

214
experiments/phase4_quality.py Normal file
View File

@@ -0,0 +1,214 @@
#!/usr/bin/env python3
"""Phase 4: End-to-End Context Quality Evaluation"""
import sys, os, json
sys.path.insert(0, os.path.dirname(os.path.dirname(__file__)))
from src.gatekeeper import ContextGatekeeper
def estimate_tokens(text):
if not text: return 0
chinese = sum(1 for c in text if '\u4e00' <= c <= '\u9fff')
english = len([w for w in text.split() if w.isascii()])
return int(chinese * 0.4 + english * 1.3 + len(text) * 0.05)
def measure_prompt_tokens(selected, query):
ctx = "".join(f"用户: {i['user']}\n助手: {i['assistant']}\n\n" for i in selected)
return estimate_tokens(ctx) + int(estimate_tokens(ctx) * 0.08) + estimate_tokens(query)
def evaluate_context_quality(gate, query, target_topic, context):
total_blocks = len(context)
topic_blocks = 0
other_topics_found = []
for item in context:
text = item['user'] + ' ' + item['assistant']
found = [t for t in ['Redis', 'asyncio', 'PostgreSQL', 'Git']
if t.lower() in text.lower() and t.lower() != target_topic.lower()]
if found:
other_topics_found.extend(found)
else:
topic_blocks += 1
block_purity = topic_blocks / total_blocks if total_blocks > 0 else 0
q_anchors, _ = gate.anchor_extractor.extract_with_deictic(query)
context_text = ' '.join(item['user'] + ' ' + item['assistant'] for item in context)
covered = sum(1 for a in q_anchors if a.lower() in context_text.lower())
anchor_recall = covered / len(q_anchors) if q_anchors else 0
key_terms = {
'PostgreSQL': ['explain', 'analyze', 'mvcc', 'vacuum', '索引', '执行计划'],
'Git': ['rebase', 'merge', 'reset', 'commit', '分支'],
'Redis': ['redis', '分布式锁', '惰性删除', '定期删除', '过期'],
'asyncio': ['asyncio', 'task', 'cancel', '事件循环', '协程'],
}
relevant_terms = key_terms.get(target_topic, [])
context_lower = context_text.lower()
terms_found = [t for t in relevant_terms if t.lower() in context_lower]
term_coverage = len(terms_found) / len(relevant_terms) if relevant_terms else 0
return {
'block_purity': block_purity,
'anchor_recall': anchor_recall,
'term_coverage': term_coverage,
'other_topics': list(set(other_topics_found)),
'context_tokens': estimate_tokens(context_text),
}
def evaluate_retrieval_quality(gate, query, target_topic):
sel_cgk = gate.select(query)
last5_context = gate.blocks[-5:]
last5_items = [{'user': b.user_text, 'assistant': b.assistant_text} for b in last5_context]
cgk_quality = evaluate_context_quality(gate, query, target_topic, sel_cgk)
last5_quality = evaluate_context_quality(gate, query, target_topic, last5_items)
cgk_prompt_tok = measure_prompt_tokens(sel_cgk, query)
last5_prompt_tok = measure_prompt_tokens(last5_items, query)
return {
'cgk': {**cgk_quality, 'prompt_tokens': cgk_prompt_tok},
'last5': {**last5_quality, 'prompt_tokens': last5_prompt_tok},
}
redis_qa = [
("Redis 分布式锁和 RedLock 算法有什么区别？", "RedLock是..."),
("Redis 集群环境下怎么做分布式锁？", "集群下..."),
("Redis 惰性删除和定期删除有什么区别？", "惰性删除..."),
("Redis 的过期 key 对 RDB 快照有什么影响？", "过期key..."),
("Redis 主从复制断线后如何增量同步？", "PSYNC..."),
]
asyncio_qa = [
("asyncio.Task 的 cancel 方法怎么工作的？", "cancel..."),
("asyncio.gather 和 asyncio.wait 的返回结果有什么区别？", "gather..."),
("asyncio 的事件循环怎么启动和停止？", "事件循环..."),
("asyncio.sleep 和 time.sleep 的区别是什么？", "sleep..."),
("asyncio 的 Future 对象怎么获取结果？", "Future..."),
]
pg_qa = [
("PostgreSQL 的 MVCC 机制是怎么保证读不阻塞写的？", "MVCC..."),
("PostgreSQL 的 VACUUM 为什么要定期运行？", "VACUUM..."),
("EXPLAIN ANALYZE 怎么看执行计划？", "EXPLAIN..."),
("PostgreSQL B-tree 索引和 Hash 索引的区别是什么？", "B-tree..."),
("PostgreSQL 的 TOAST 机制是什么？", "TOAST..."),
]
git_qa = [
("Git 的 rebase 和 merge 的区别是什么？", "rebase..."),
("Git reset 的 --soft、--mixed、--hard 有什么区别？", "reset..."),
("Git stash 暂存区和工作目录的区别是什么？", "stash..."),
("Git 的 bisect 怎么用来快速定位 bug？", "bisect..."),
("Git 的 reflog 怎么用来恢复误删的提交？", "reflog..."),
]
TEST_SEQ = [
("问PG", "EXPLAIN ANALYZE 怎么看执行计划？", "PostgreSQL"),
("问Git", "Git 的 rebase 和 merge 有什么区别？", "Git"),
("问Redis", "Redis 惰性删除和定期删除有什么区别？", "Redis"),
("问asyncio", "asyncio.Task 的 cancel 方法怎么工作的？", "asyncio"),
("再问Git", "Git 的 reset 和 revert 的应用场景有什么区别？", "Git"),
]
def build_gate():
g = ContextGatekeeper(token_budget=4000)
for i in range(5):
g.add_turn(redis_qa[i][0], redis_qa[i][1])
g.add_turn(asyncio_qa[i][0], asyncio_qa[i][1])
g.add_turn(pg_qa[i][0], pg_qa[i][1])
g.add_turn(git_qa[i][0], git_qa[i][1])
return g
def main():
print("="*70)
print("Phase 4: End-to-End Context Quality Evaluation")
print("="*70)
print("评估维度:")
print(" - block_purity: 目标话题块占比（1.0=纯目标话题）")
print(" - anchor_recall: query锚点在上下文中的覆盖率")
print(" - term_coverage: 目标话题关键术语在上下文中的覆盖率")
print(" - prompt_tokens: 完整prompt token数（含格式化开销）")
print()
cgk_tokens_list, last5_tokens_list = [], []
cgk_purities, last5_purities = [], []
cgk_anchors, last5_anchors = [], []
cgk_terms, last5_terms = [], []
cgk_cont, last5_cont = [], []
for label, query, target in TEST_SEQ:
gate = build_gate()
r = evaluate_retrieval_quality(gate, query, target)
cgk = r['cgk']
last5 = r['last5']
cgk_tokens_list.append(cgk['prompt_tokens'])
last5_tokens_list.append(last5['prompt_tokens'])
cgk_purities.append(cgk['block_purity'])
last5_purities.append(last5['block_purity'])
cgk_anchors.append(cgk['anchor_recall'])
last5_anchors.append(last5['anchor_recall'])
cgk_terms.append(cgk['term_coverage'])
last5_terms.append(last5['term_coverage'])
cgk_cont.append(1 if cgk['other_topics'] else 0)
last5_cont.append(1 if last5['other_topics'] else 0)
print(f"\n[{label}] {query}")
print(f" CGK: tok={cgk['prompt_tokens']:.0f}, purity={cgk['block_purity']:.2f}, "
f"anchor={cgk['anchor_recall']:.2f}, term={cgk['term_coverage']:.2f}, "
f"other_topics={cgk['other_topics']}")
print(f" Last-5: tok={last5['prompt_tokens']:.0f}, purity={last5['block_purity']:.2f}, "
f"anchor={last5['anchor_recall']:.2f}, term={last5['term_coverage']:.2f}, "
f"other_topics={last5['other_topics']}")
saving = (1 - cgk['prompt_tokens']/last5['prompt_tokens'])*100
print(f" → CGK节省{saving:.0f}%token, 纯度+{(cgk['block_purity']-last5['block_purity'])*100:.0f}%, "
f"混入话题{len(cgk['other_topics'])}个 vs Last-5 {len(last5['other_topics'])}个")
n = len(TEST_SEQ)
print("\n" + "="*70)
print("Summary: CGK vs Last-5 (avg over {} queries)".format(n))
print("="*70)
print(f"\n{'Metric':<25} {'CGK':>12} {'Last-5':>12} {'Winner':<10}")
print("-"*62)
cgk_avg_tok = sum(cgk_tokens_list)/n
last5_avg_tok = sum(last5_tokens_list)/n
print(f"{'Avg prompt tokens':<25} {cgk_avg_tok:>12.1f} {last5_avg_tok:>12.1f} {'CGK' if cgk_avg_tok < last5_avg_tok else 'Last-5':<10}")
cgk_avg_pur = sum(cgk_purities)/n
last5_avg_pur = sum(last5_purities)/n
print(f"{'Avg block purity':<25} {cgk_avg_pur:>12.3f} {last5_avg_pur:>12.3f} {'CGK' if cgk_avg_pur > last5_avg_pur else 'Last-5':<10}")
cgk_avg_anc = sum(cgk_anchors)/n
last5_avg_anc = sum(last5_anchors)/n
print(f"{'Avg anchor recall':<25} {cgk_avg_anc:>12.3f} {last5_avg_anc:>12.3f} {'CGK' if cgk_avg_anc > last5_avg_anc else 'Last-5':<10}")
cgk_avg_term = sum(cgk_terms)/n
last5_avg_term = sum(last5_terms)/n
print(f"{'Avg term coverage':<25} {cgk_avg_term:>12.3f} {last5_avg_term:>12.3f} {'CGK' if cgk_avg_term > last5_avg_term else 'Last-5':<10}")
print(f"{'Contamination episodes':<25} {sum(cgk_cont):>12} {sum(last5_cont):>12} {'CGK' if sum(cgk_cont) < sum(last5_cont) else 'Last-5':<10}")
print("\n" + "="*70)
print("Honest Limitations of This Evaluation")
print("="*70)
print("""
1. 没有真实 LLM 调用：评估的是"上下文块的质量"，不是"模型答案的质量"。
上下文好 ≠ 答案好，真正的答案质量需要实际调用 LLM。
2. 测试集仍是合成数据：真实对话中用户可能只打"那这个呢"或"为什么"，
短 query 的锚点覆盖率会显著低于本测试中的完整问题。
3. 污染只统计了"块级别"混入：即使上下文纯度 100%，LLM 的注意力机制
仍可能跨块建立错误关联，这种"软污染"无法通过块级分析检测。
""")
out_path = os.path.join(os.path.dirname(__file__), 'phase4_quality_results.json')
with open(out_path, 'w') as f:
json.dump({
'cgk_avg_tokens': cgk_avg_tok,
'last5_avg_tokens': last5_avg_tok,
'cgk_avg_purity': cgk_avg_pur,
'last5_avg_purity': last5_avg_pur,
'cgk_avg_anchor_recall': cgk_avg_anc,
'last5_avg_anchor_recall': last5_avg_anc,
'cgk_avg_term_coverage': cgk_avg_term,
'last5_avg_term_coverage': last5_avg_term,
'cgk_contamination_episodes': sum(cgk_cont),
'last5_contamination_episodes': sum(last5_cont),
}, f, indent=2)
print(f"\nSaved to: {out_path}")
if __name__ == '__main__':
main()

12
experiments/phase4_quality_results.json Normal file
View File

@@ -0,0 +1,12 @@
{
"cgk_avg_tokens": 42.2,
"last5_avg_tokens": 67.6,
"cgk_avg_purity": 1.0,
"last5_avg_purity": 0.28,
"cgk_avg_anchor_recall": 0.6382417582417583,
"last5_avg_anchor_recall": 0.06598502946329034,
"cgk_avg_term_coverage": 0.38,
"last5_avg_term_coverage": 0.08,
"cgk_contamination_episodes": 0,
"last5_contamination_episodes": 5
}