fix: agent search uses smart_search for content queries

- Update system prompt to prioritize smart_search over find_file
- Add source prefix ([OCR], [ASRX], [ASRX+OCR]) to exec_smart_search results
- Users can now see chunk content with source indicators
This commit is contained in:
Accusys
2026-07-19 14:21:12 +08:00
parent 87aa7e0c40
commit 3c924e1f83
2 changed files with 87 additions and 17 deletions
+71 -5
View File
@@ -555,7 +555,7 @@ pub async fn exec_smart_search(
let chunk_table = schema::table_name("chunk");
let mut sql = format!(
"SELECT chunk_id, text_content, start_frame, end_frame, chunk_type \
"SELECT chunk_id, text_content, start_frame, end_frame, chunk_type, content \
FROM {} WHERE text_content ILIKE $1",
chunk_table
);
@@ -566,21 +566,87 @@ pub async fn exec_smart_search(
if let Some(fuid) = file_uuid {
let like = format!("%{}%", query);
let rows: Vec<(String, Option<String>, i64, i64, String)> = sqlx::query_as(&sql)
let rows: Vec<(String, Option<String>, i64, i64, String, Option<serde_json::Value>)> = sqlx::query_as(&sql)
.bind(&like)
.bind(fuid)
.fetch_all(pool)
.await
.map_err(|e| e.to_string())?;
Ok(serde_json::json!({"results": rows}).to_string())
let results: Vec<Vec<serde_json::Value>> = rows.into_iter().map(|(chunk_id, text_content, start_frame, end_frame, chunk_type, content)| {
let source_prefix = if let Some(ref content) = content {
let text = content.get("text").and_then(|t| t.as_str()).unwrap_or("");
let ocr_text = content.get("ocr_text").and_then(|t| t.as_str()).unwrap_or("");
let has_asrx = !text.trim().is_empty();
let has_ocr = !ocr_text.trim().is_empty();
if has_asrx && has_ocr {
"[ASRX+OCR] "
} else if has_asrx {
"[ASRX] "
} else if has_ocr {
"[OCR] "
} else {
""
}
} else {
""
};
let prefixed_text = text_content.map(|t| format!("{}{}", source_prefix, t));
vec![
serde_json::json!(chunk_id),
serde_json::json!(prefixed_text),
serde_json::json!(start_frame),
serde_json::json!(end_frame),
serde_json::json!(chunk_type),
]
}).collect();
Ok(serde_json::json!({"results": results}).to_string())
} else {
let like = format!("%{}%", query);
let rows: Vec<(String, Option<String>, i64, i64, String)> = sqlx::query_as(&sql)
let rows: Vec<(String, Option<String>, i64, i64, String, Option<serde_json::Value>)> = sqlx::query_as(&sql)
.bind(&like)
.fetch_all(pool)
.await
.map_err(|e| e.to_string())?;
Ok(serde_json::json!({"results": rows}).to_string())
let results: Vec<Vec<serde_json::Value>> = rows.into_iter().map(|(chunk_id, text_content, start_frame, end_frame, chunk_type, content)| {
let source_prefix = if let Some(ref content) = content {
let text = content.get("text").and_then(|t| t.as_str()).unwrap_or("");
let ocr_text = content.get("ocr_text").and_then(|t| t.as_str()).unwrap_or("");
let has_asrx = !text.trim().is_empty();
let has_ocr = !ocr_text.trim().is_empty();
if has_asrx && has_ocr {
"[ASRX+OCR] "
} else if has_asrx {
"[ASRX] "
} else if has_ocr {
"[OCR] "
} else {
""
}
} else {
""
};
let prefixed_text = text_content.map(|t| format!("{}{}", source_prefix, t));
vec![
serde_json::json!(chunk_id),
serde_json::json!(prefixed_text),
serde_json::json!(start_frame),
serde_json::json!(end_frame),
serde_json::json!(chunk_type),
]
}).collect();
Ok(serde_json::json!({"results": results}).to_string())
}
}