fix(memory): CJK-aware short query threshold + Chinese synonym expansion
1. MemoryMiddleware: replace byte-length check (query.len() < 4) with char-count check (query.chars().count() < 2). Single CJK characters are 3 UTF-8 bytes but 1 meaningful character — the old threshold incorrectly skipped 1-2 char Chinese queries like "你好". 2. QueryAnalyzer: add Chinese synonym mappings for 13 common technical terms (错误→bug, 优化→improve, 配置→config, etc.) so CJK queries can find relevant English-keyword memories and vice versa. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
@@ -216,9 +216,10 @@ impl QueryAnalyzer {
|
||||
expansions
|
||||
}
|
||||
|
||||
/// Get synonyms for a keyword (simplified)
|
||||
/// Get synonyms for a keyword (simplified, English + Chinese)
|
||||
fn get_synonyms(&self, keyword: &str) -> Option<Vec<String>> {
|
||||
let synonyms: &[&str] = match keyword {
|
||||
// English synonyms
|
||||
"code" => &["program", "script", "source"],
|
||||
"error" => &["bug", "issue", "problem", "exception"],
|
||||
"fix" => &["solve", "resolve", "repair", "patch"],
|
||||
@@ -226,6 +227,20 @@ impl QueryAnalyzer {
|
||||
"slow" => &["performance", "optimize", "speed"],
|
||||
"help" => &["assist", "support", "guide", "aid"],
|
||||
"learn" => &["study", "understand", "know", "grasp"],
|
||||
// Chinese synonyms — critical for Chinese-language queries
|
||||
"错误" => &["问题", "bug", "异常", "故障"],
|
||||
"修复" => &["解决", "修正", "处理", "fix"],
|
||||
"优化" => &["改进", "提升", "加速", "improve"],
|
||||
"配置" => &["设置", "参数", "选项", "config"],
|
||||
"性能" => &["速度", "效率", "performance"],
|
||||
"问题" => &["错误", "故障", "issue", "problem"],
|
||||
"帮助" => &["协助", "支持", "help"],
|
||||
"学习" => &["了解", "掌握", "learn"],
|
||||
"代码" => &["程序", "脚本", "code"],
|
||||
"数据库" => &["DB", "database", "存储"],
|
||||
"部署" => &["发布", "上线", "deploy"],
|
||||
"测试" => &["验证", "检验", "test"],
|
||||
"安全" => &["防护", "加密", "security"],
|
||||
_ => return None,
|
||||
};
|
||||
|
||||
|
||||
@@ -60,6 +60,22 @@ impl AgentMiddleware for MemoryMiddleware {
|
||||
fn priority(&self) -> i32 { 150 }
|
||||
|
||||
async fn before_completion(&self, ctx: &mut MiddlewareContext) -> Result<MiddlewareDecision> {
|
||||
// Skip memory injection for very short queries.
|
||||
// Short queries (e.g., "1+6", "hi", "好") don't benefit from memory context.
|
||||
// Worse, the retriever's scope-based fallback may return high-importance but
|
||||
// irrelevant old memories, causing the model to think about past conversations
|
||||
// instead of answering the current question.
|
||||
// Use char count (not byte count) so CJK queries are handled correctly:
|
||||
// a single Chinese char is 3 UTF-8 bytes but 1 meaningful character.
|
||||
let query = ctx.user_input.trim();
|
||||
if query.chars().count() < 2 {
|
||||
tracing::debug!(
|
||||
"[MemoryMiddleware] Skipping enhancement for short query ({:?}): no memory context needed",
|
||||
query
|
||||
);
|
||||
return Ok(MiddlewareDecision::Continue);
|
||||
}
|
||||
|
||||
match self.growth.enhance_prompt(
|
||||
&ctx.agent_id,
|
||||
&ctx.system_prompt,
|
||||
@@ -92,21 +108,27 @@ impl AgentMiddleware for MemoryMiddleware {
|
||||
return Ok(());
|
||||
}
|
||||
|
||||
match self.growth.process_conversation(
|
||||
// Combined extraction: single LLM call produces both memories and structured facts.
|
||||
// Avoids double LLM extraction ( process_conversation + extract_structured_facts).
|
||||
match self.growth.extract_combined(
|
||||
&ctx.agent_id,
|
||||
&ctx.messages,
|
||||
ctx.session_id.clone(),
|
||||
&ctx.session_id,
|
||||
).await {
|
||||
Ok(count) => {
|
||||
Ok(Some((mem_count, facts))) => {
|
||||
tracing::info!(
|
||||
"[MemoryMiddleware] Extracted {} memories for agent {}",
|
||||
count,
|
||||
"[MemoryMiddleware] Extracted {} memories + {} structured facts for agent {}",
|
||||
mem_count,
|
||||
facts.len(),
|
||||
agent_key
|
||||
);
|
||||
}
|
||||
Ok(None) => {
|
||||
tracing::debug!("[MemoryMiddleware] No memories or facts extracted");
|
||||
}
|
||||
Err(e) => {
|
||||
// Non-fatal: extraction failure should not affect the response
|
||||
tracing::warn!("[MemoryMiddleware] Memory extraction failed: {}", e);
|
||||
tracing::warn!("[MemoryMiddleware] Combined extraction failed: {}", e);
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
Reference in New Issue
Block a user