fix: improve search accuracy with multi-token LIKE and guided agent search queries

- _like_query: split multi-token queries into per-field AND matching
- lookup_paper: same multi-token AND for title lookup
- read-known-paper.md: restructure Step 1 by known-info type (zotero_key,
  DOI, author+year, title keyword, fuzzy)
- discover-papers.md Arm 2: note on author+year search pattern
- retrieval-routing.md Arm 2: same author+year search guidance

paperforge/memory/fts.py

@@ -82,7 +82,18 @@ def _fts_query(conn, query, filter_clause, filter_params, limit):
 
 
 def _like_query(conn, query, filter_clause, filter_params, limit):
-    like_param = f"%{query}%"
+    tokens = re.findall(r"[\w\u4e00-\u9fff]+", query)
+    if not tokens:
+        tokens = [query]
+
+    field_clauses = []
+    params = []
+    for field in ["p.title", "p.abstract", "p.doi", "p.citation_key"]:
+        token_clauses = [f"{field} LIKE ?" for _ in tokens]
+        params.extend(f"%{t}%" for t in tokens)
+        field_clauses.append("(" + " AND ".join(token_clauses) + ")")
+
+    where_clause = "(" + " OR ".join(field_clauses) + ")"
     sql = f"""
         SELECT p.zotero_key, p.citation_key, p.title, p.year, p.doi,
                p.first_author, p.journal, p.domain, p.lifecycle,
@@ -90,10 +101,10 @@ def _like_query(conn, query, filter_clause, filter_params, limit):
                substr(p.abstract, 1, 300) as abstract,
                0 as rank
         FROM papers p
-        WHERE (p.title LIKE ? OR p.abstract LIKE ? OR p.doi LIKE ? OR p.citation_key LIKE ?){filter_clause}
+        WHERE {where_clause}{filter_clause}
         ORDER BY p.year DESC
         LIMIT ?
     """
     conn.row_factory = sqlite3.Row
-    rows = conn.execute(sql, [like_param, like_param, like_param, like_param] + filter_params + [limit]).fetchall()
+    rows = conn.execute(sql, params + filter_params + [limit]).fetchall()
     return [dict(r) for r in rows]

paperforge/memory/query.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import json
 import logging
+import re
 from pathlib import Path
 
 from paperforge.memory.builder import compute_hash
@@ -79,7 +80,7 @@ def get_memory_status(vault: Path) -> dict:
 
 def _entry_from_row(row) -> dict:
     """Reconstruct an entry dict from a papers row (sqlite3.Row)."""
-    entry = {k: row[k] for k in row.keys()}
+    entry = {k: row[k] for k in row}
     for key in ("has_pdf", "do_ocr", "analyze"):
         if key in entry and entry[key] is not None:
             entry[key] = bool(entry[key])
@@ -108,11 +109,16 @@ def lookup_paper(conn, query: str) -> list[dict]:
         if row:
             return [_entry_from_row(row)]
 
+    tokens = re.findall(r"[\w\u4e00-\u9fff]+", q)
+    if not tokens:
+        tokens = [q]
+    title_conditions = " AND ".join("LOWER(title) LIKE ?" for _ in tokens)
+    title_params = [f"%{t.lower()}%" for t in tokens]
     rows = conn.execute(
-        """SELECT * FROM papers
-           WHERE LOWER(title) LIKE '%' || LOWER(?) || '%'
+        f"""SELECT * FROM papers
+           WHERE {title_conditions}
            LIMIT 20""",
-        (q,),
+        title_params,
     ).fetchall()
     if rows:
         return [_entry_from_row(r) for r in rows]

paperforge/skills/paperforge/atoms/retrieval-routing.md

@@ -56,6 +56,7 @@ $PYTHON -m paperforge --vault "$VAULT" search "<query>" --json --limit 30
 - Uses FTS5 on title, abstract, authors, journal, domain, collection_path
 - Fast, always works
 - Limit 30 by default; user can request more
+- **Author+year search**: put author name in `<query>`, year in `--year-from`/`--year-to` (e.g. `search "Brighton" --year-from 1983 --year-to 1983`)
 
 #### Arm 3 — Collection/domain inventory (full list, no truncation)
 

paperforge/skills/paperforge/molecules/discover-papers.md

@@ -68,6 +68,7 @@ $PYTHON -m paperforge --vault "$VAULT" search "<query>" --json --limit 30 \
 
 - FTS5 搜索标题、摘要、作者、期刊、domain、collection 路径
 - 返回 JSON：`data.matches[]` 包含 `zotero_key`、`title`、`year`、`first_author` 等
+- **作者+年份搜索**：query 放作者名，年份放 `--year-from`/`--year-to`（如 `search "Brighton" --year-from 1983 --year-to 1983`）
 
 #### Arm 3 — Collection/Domain 列举（完整列表，不截断）
 

paperforge/skills/paperforge/molecules/read-known-paper.md

@@ -25,31 +25,46 @@
 
 ### Step 1: 定位论文
 
-用户可能给 zotero_key、DOI、标题片段、作者+年份。按以下方式查找：
+用户可能给 zotero_key、DOI、标题关键词、作者+年份。按以下方式查找：
 
-**如果用户给的是 zotero_key：**
+**如果知道 zotero_key：**
 
 ```bash
 $PYTHON -m paperforge --vault "$VAULT" paper-context <zotero_key> --json
 ```
 
-返回 JSON 包含 paper 元数据、OCR 状态、prior_notes 等。
+**如果知道 DOI：**
 
-**如果用户给的是 DOI、标题片段、作者+年份，先解析成候选：**
+```bash
+$PYTHON -m paperforge --vault "$VAULT" paper-context <DOI> --json
+```
+
+**如果知道 作者+年份：**
+
+```bash
+$PYTHON -m paperforge --vault "$VAULT" search "<author>" --year-from <YYYY> --year-to <YYYY> --json --limit 10
+```
+
+**如果知道 标题关键词：**
+
+```bash
+$PYTHON -m paperforge --vault "$VAULT" search "<关键词>" --json --limit 10
+```
+
+**如果只有模糊描述：**
 
 ```bash
 $PYTHON -m paperforge --vault "$VAULT" paper-status "<query>" --json
 ```
 
-如果 `paper-status` 直接唯一命中，取返回的 `zotero_key` 再调 `paper-context`。
-
-**paper-status 无法唯一定位时的备选：**
+如果无唯一命中：
 
 ```bash
 $PYTHON -m paperforge --vault "$VAULT" search "<query>" --json --limit 5
 ```
 
-如果多候选，列出让用户选；用户选定后，再用选中的 `zotero_key` 调 `paper-context`。
+无论哪种方式命中，取 `zotero_key` 调 `paper-context` 获取完整信息。
+如果多候选，列出让用户选。
 如果无结果，告知用户并停止。
 
 ### Step 2: 加载文献内容