import json import tempfile import time from pathlib import Path import gzip import hashlib from src.adapters.indexes.knowledge_sqlite import SqliteKnowledgeRepository from src.modules.knowledge.application.catalog import KnowledgeCatalog from src.modules.skill.domain.models import Actor from src.modules.knowledge.tests.test_knowledge import write_fixture BUDGET_BYTES = 8000 QUERY = "Java ERP 采购订单审批" actor = Actor( tenant_id="test-tenant", user_id="developer", roles=[], ) def size_bytes(record): return len(str(record).encode("utf-8")) def original_break(records): selected = [] used = 0 for record in records: size = size_bytes(record) if used + size > BUDGET_BYTES: break selected.append(record) used += size return selected def improved_continue(records): selected = [] used = 0 for record in records: size = size_bytes(record) if used + size > BUDGET_BYTES: continue selected.append(record) used += size return selected def top2(records): return improved_continue(records[:2]) def measure(name, fn, records): start = time.perf_counter() selected = fn(records) elapsed_ms = (time.perf_counter() - start) * 1000 print(f"\n[{name}]") print("保留文档:", [r["citation_id"] for r in selected]) print("总字节:", sum(size_bytes(r) for r in selected)) print("处理耗时(ms):", round(elapsed_ms, 4)) def main(): with tempfile.TemporaryDirectory() as temp: root = Path(temp) repo = SqliteKnowledgeRepository(root / "index.sqlite") corpus_dir = root / "corpus" catalog_path = write_benchmark_corpus(corpus_dir) repo.import_corpus( corpus_dir, actor.tenant_id, catalog_path, ) catalog = KnowledgeCatalog(repo) start = time.perf_counter() records = catalog.search( actor, QUERY, ["fixture-java"], limit=8, ) retrieval_ms = (time.perf_counter() - start) * 1000 print("KnowledgeCatalog 检索文档数:", len(records)) print("检索耗时(ms):", round(retrieval_ms, 3)) measure("A 原版 break", original_break, records) measure("B 改进 continue", improved_continue, records) measure("C Top-2", top2, records) if len(records) < 3: print( "\n注意:测试库文档不足,暂时无法有效比较筛选效果。" ) def write_benchmark_corpus(folder): """生成 8 条用于测试检索和预算筛选的 Java 企业知识。""" folder.mkdir(parents=True, exist_ok=True) specs = [ ("java-style", "Java 编码规范", "Java 类名使用 PascalCase,方法名使用 camelCase。"), ("erp-order", "ERP 采购订单流程", "采购订单应包含供应商、商品明细、数量和金额。"), ("wms-stock", "WMS 库存流水", "库存流水记录商品入库、出库及剩余库存数量。"), ("erp-approval", "ERP 采购审批核心规则", "ERP 采购订单提交后必须经过审批,审批通过后才能执行采购操作。"), ("spring-api", "Spring Boot REST API", "Spring Boot 使用 REST Controller 实现 Java HTTP API。"), ("transaction", "Java 数据库事务", "Java 事务处理需要保证提交和回滚的一致性。"), ("maintenance", "设备维修管理", "设备维修管理记录工单、检修状态和故障信息。"), ("java-test", "Java 单元测试规范", "Java 单元测试通过 JUnit 验证接口行为。"), ] records = [] for index, (record_id, title, content) in enumerate(specs): content = content * (40 if record_id == "erp-order" else 5) record = { "record_id": record_id, "collection_id": "fixture-java", "title": title, "content": content, "content_sha256": hashlib.sha256(content.encode()).hexdigest(), "source_url": f"https://example.com/{record_id}", "source_path": f"{record_id}.java", "source_commit": "a" * 40, "line_start": 1, "line_end": 10, "license": "Apache-2.0", "usage_policy": "attribution-required", "quality_status": "imported_unreviewed", "trust": "external_reference", } records.append(record) with gzip.open( folder / "fixture-java.jsonl.gz", "wt", encoding="utf-8" ) as f: for record in records: f.write(json.dumps(record, ensure_ascii=False) + "\n") catalog = folder / "catalog.json" # 生成符合原项目导入格式的 manifest.json data_file = folder / "fixture-java.jsonl.gz" manifest = { "collections": [ { "id": "fixture-java", "name": "Java ERP 测试知识库", "description": "Token optimization benchmark", "sha256": hashlib.sha256( data_file.read_bytes() ).hexdigest(), "records": len(records), "encoding": "utf-8", } ] } (folder / "manifest.json").write_text( json.dumps(manifest, ensure_ascii=False), encoding="utf-8", ) catalog.write_text( json.dumps( {"collections": [{"id": "fixture-java"}]}, ensure_ascii=False, ), encoding="utf-8", ) return catalog if __name__ == "__main__": main()