From 5e83ee7dac740064d28187657a0208e166f32635 Mon Sep 17 00:00:00 2001
From: Accusys <accusys@momentry.dev>
Date: Sun, 19 Jul 2026 13:33:31 +0800
Subject: [PATCH] fix: keyword search for OCR text
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- Fix OCR confidence threshold: 0.5 → 0.2
- Fix fetch_ocr_texts frame calculation (use start_frame from DB)
- Fix text_match filter to use text_content instead of summary
- Process all merged results (not just top 30) to include keyword results
- Add logging for keyword search debugging

Fixes issue where OCR text like 'AUDIO MONITORING' and 'thunderbolt'
could not be found via keyword search.
---
 src/api/search.rs              | 38 +++++++++++++++++++++++++++-------
 src/core/chunk/rule1_ingest.rs |  5 ++---
 2 files changed, 33 insertions(+), 10 deletions(-)

diff --git a/src/api/search.rs b/src/api/search.rs
index d9abbbe..095e660 100644
--- a/src/api/search.rs
+++ b/src/api/search.rs
@@ -191,10 +191,16 @@ pub async fn smart_search(
         .search_bm25(&req.query, req.file_uuid.as_deref(), fetch_limit as i64)
         .await
     {
-        Ok(rows) => rows
-            .into_iter()
-            .map(|r| (r.file_uuid, r.chunk_id, r.combined_score))
-            .collect(),
+        Ok(rows) => {
+            tracing::info!(
+                "Smart search: Keyword search returned {} hits for '{}'",
+                rows.len(),
+                req.query
+            );
+            rows.into_iter()
+                .map(|r| (r.file_uuid, r.chunk_id, r.combined_score))
+                .collect()
+        }
         Err(e) => {
             tracing::warn!("Keyword search (bm25) failed: {}", e);
             vec![]
@@ -379,10 +385,22 @@ pub async fn smart_search(
             .unwrap_or(std::cmp::Ordering::Equal)
     });
 
+    tracing::info!(
+        "Smart search: Merged {} results (semantic={}, keyword={}, title={}, identity={})",
+        merged.len(),
+        semantic_results.len(),
+        keyword_results.len(),
+        title_results.len(),
+        identity_results.len()
+    );
+
     // 6. Enrich top results from PG and build final response
     let query_lower = req.query.to_lowercase();
     let mut final_results = Vec::new();
-    for mr in ranked.iter().take(limit * 3) {
+    // Take all merged results to ensure keyword results are processed
+    // (keyword results often have low scores and rank last)
+    let take_count = ranked.len();
+    for mr in ranked.iter().take(take_count) {
         // 取更多結果以便過濾
         // Use no_embedding version for keyword results, regular for semantic
         let pg_opt = if mr.keyword_score.is_some() && mr.semantic_score.is_none() {
@@ -397,13 +415,19 @@ pub async fn smart_search(
             let is_keyword_only = mr.keyword_score.is_some() && mr.semantic_score.is_none();
             if !is_keyword_only {
                 // 關鍵字過濾: CJK 用子字串匹配，英文用單詞邊界匹配
-                let summary_lower = pg.summary.to_lowercase();
+                // 使用 text_content 或 summary 進行匹配
+                let match_text = if pg.summary.is_empty() {
+                    pg.text_content.clone().unwrap_or_default()
+                } else {
+                    pg.summary.clone()
+                };
+                let summary_lower = match_text.to_lowercase();
                 let query_words: Vec<String> = query_lower
                     .split_whitespace()
                     .map(|s| s.to_string())
                     .collect();
 
-                let text_match = !pg.summary.is_empty() && {
+                let text_match = !match_text.is_empty() && {
                     let has_cjk = |s: &str| -> bool {
                         s.chars().any(|c| {
                             ('\u{4E00}'..='\u{9FFF}').contains(&c)
diff --git a/src/core/chunk/rule1_ingest.rs b/src/core/chunk/rule1_ingest.rs
index 847b499..3117cf5 100644
--- a/src/core/chunk/rule1_ingest.rs
+++ b/src/core/chunk/rule1_ingest.rs
@@ -7,7 +7,7 @@ use sqlx::{PgPool, Row};
 use std::collections::BTreeMap;
 use tracing::{info, warn};
 
-const OCR_CONFIDENCE_THRESHOLD: f64 = 0.5;
+const OCR_CONFIDENCE_THRESHOLD: f64 = 0.2;
 
 pub async fn execute_rule1(db: &PostgresDb, file_uuid: &str, fps: f64) -> Result<usize> {
     let pool = db.pool();
@@ -219,11 +219,10 @@ async fn fetch_ocr_texts(
 
     let mut map: BTreeMap<i64, Vec<String>> = BTreeMap::new();
     for row in rows {
-        let start_time: f64 = row.try_get("start_time").unwrap_or(0.0);
+        let start_frame: i64 = row.try_get("start_frame").unwrap_or(0);
         let end_time_raw: Option<f64> = row.try_get("end_time").ok();
         let data: Value = row.try_get("data").unwrap_or(Value::Null);
 
-        let start_frame = (start_time * fps) as i64;
         let end_frame = end_time_raw
             .filter(|t| *t > 0.0)
             .map(|t| (t * fps) as i64)