Unverified50% confidenceBenchmarkExact time
In the empirical tests, the MiniLM Cross-Encoder, biased toward lexical matching and sensitive to the scrambled text sequences produced by PDF parsing, pushed both correct answer chunks out of the Top 5, causing ranking degradation.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/14/2026
First Seen
Valid until: 1/12/2027
Sources
Related Entities
Related Claims
UnverifiedFTS5 依赖字面匹配,当查询词与存储内容用词不同但语义相近时可能漏检,是纯全文方案的固有局限72% similarUnverifiedBag-of-words cosine similarity splits text into a set of tokens, ignoring order and counting only frequency; it is sensitive to vocabulary overlap but insensitive to word order.70% similarUnverifiedUsing a simple text similarity threshold or difference detection can easily misclassify reasonable wording changes as violations.68% similarUnverifiedWhen complex PDFs are naively extracted into plain text, their inherent hierarchical relationships and semantic boundaries are destroyed, affecting the accuracy of subsequent retrieval.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1028161API
curl https://kongchang.com/api/v1/knowledge/claims/1028161MCP
get_claim(id=1028161)