Skip to content

Codebase intelligence for AI and humans: code health scores, auto-generated docs, git analytics, dead code detection, and architectural decisions via MCP.

fix(search): match path queries against filename words (#3180)

Raghav Chamadiya82f4ebb2aaaf
diff --git a/packages/server/src/repowise/server/mcp_server/_retrieval_rank.py b/packages/server/src/repowise/server/mcp_server/_retrieval_rank.py
index 921299e9f..afc798e24 100644
--- a/packages/server/src/repowise/server/mcp_server/_retrieval_rank.py
+++ b/packages/server/src/repowise/server/mcp_server/_retrieval_rank.py
@@ -159,10 +159,10 @@ _PATH_NAME_LIFT_CAP = 1.0
_MIN_INDEXED_PATHS = 30
-def path_words(path: str) -> set[str]:
+def path_words(path: str, *, min_len: int = 3) -> set[str]:
"""``services/_hotkey_pynput.py`` -> ``{services, hotkey, pynput}``; the
extension is not a word."""
- return {w for w in _identifier_words(os.path.splitext(path)[0]) if len(w) >= 3}
+ return {w for w in _identifier_words(os.path.splitext(path)[0]) if len(w) >= min_len}
def path_word_counts(paths: list[str]) -> Counter[str]:
diff --git a/packages/server/src/repowise/server/mcp_server/tool_search_symbols.py b/packages/server/src/repowise/server/mcp_server/tool_search_symbols.py
index 339ec1aa5..8c7c30fc4 100644
--- a/packages/server/src/repowise/server/mcp_server/tool_search_symbols.py
+++ b/packages/server/src/repowise/server/mcp_server/tool_search_symbols.py
@@ -20,6 +20,7 @@ from __future__ import annotations
import re
from collections import Counter
+from collections.abc import Sequence
from typing import Any, NamedTuple
from sqlalchemy import case, func, or_, select
@@ -43,7 +44,11 @@ from repowise.server.mcp_server._query_shape import (
is_issue_shaped,
path_tokens,
)
-from repowise.server.mcp_server._retrieval_rank import path_word_counts
+from repowise.server.mcp_server._retrieval_rank import (
+ _identifier_words,
+ path_word_counts,
+ path_words,
+)
from repowise.server.mcp_server._stack_trace import (
frame_basenames,
map_to_repo_paths,
@@ -420,6 +425,39 @@ def _path_score(target_path: str, qnorm: str) -> float:
return score
+def _word_path_score(path: str, words: list[str]) -> float | None:
+ """Score ``path`` when every query word names one of its words (the last
+ may be a prefix), else ``None``. Below any substring hit: filename words
+ count above directory words, shorter paths first."""
+ low = path.lower()
+ if not all(w in low for w in words):
+ return None
+ every = path_words(path, min_len=1)
+ base = path_words(path.rsplit("/", 1)[-1], min_len=1)
+
+ def names(i: int, pool: set[str]) -> bool:
+ w = words[i]
+ return w in pool or (i == len(words) - 1 and any(p.startswith(w) for p in pool))
+
+ if not all(names(i, every) for i in range(len(words))):
+ return None
+ in_base = sum(names(i, base) for i in range(len(words)))
+ return 5.0 + 5.0 * in_base / len(words) - len(low) * 0.001
+
+
+def _word_path_hits(query: str, paths: Sequence[str]) -> dict[str, float]:
+ """Paths a spaced or compound query names word by word (``poll backoff``
+ -> ``command-poll-backoff.ts``), best ``_MAX_CANDIDATES`` by score. Glob
+ queries are left to substring matching."""
+ if "*" in query or "?" in query:
+ return {}
+ words = _identifier_words(query)
+ if not words:
+ return {}
+ scored = {p: s for p in paths if (s := _word_path_score(p, words)) is not None}
+ return dict(sorted(scored.items(), key=lambda kv: (-kv[1], kv[0]))[:_MAX_CANDIDATES])
+
+
class PathIndex(NamedTuple):
paths: tuple[str, ...] # every file page's path
word_counts: Counter[str] # path word -> how many paths carry it
@@ -445,7 +483,7 @@ async def search_paths_single(ctx: Any, query: str, limit: int) -> list[dict]:
Only the path-shaped words of ``query`` are matched (``path_tokens``), so
words around a path do not empty the result; the whole query is matched
- when it carries none.
+ when it carries none, as a substring and then word by word.
"""
# Path mode is substring matching, so boundary ``*?`` markers carry no
# information and are stripped. Mid-string globs (``src/*/main.py``) are
@@ -454,10 +492,16 @@ async def search_paths_single(ctx: Any, query: str, limit: int) -> list[dict]:
async with get_session(ctx.session_factory) as session:
repository = await _get_repo(session)
index = await file_path_index(session, repository.id)
- tokens = path_tokens(query, index.paths) or [query]
- qnorms = [q for q in (t.strip().lower().replace("\\", "/").strip("*?") for t in tokens) if q]
+ tokens = path_tokens(query, index.paths)
+ qnorms = [
+ q
+ for q in (t.strip().lower().replace("\\", "/").strip("*?") for t in tokens or [query])
+ if q
+ ]
if not qnorms:
return []
+ # A query carrying no path also matches paths by their words.
+ word_hits = {} if tokens else _word_path_hits(query, index.paths)
res = await session.execute(
select(Page.id, Page.title, Page.target_path, Page.freshness_status).where(
Page.repository_id == repository.id,
@@ -466,7 +510,8 @@ async def search_paths_single(ctx: Any, query: str, limit: int) -> list[dict]:
*(
Page.target_path.ilike(f"%{escape_like(q)}%", escape=LIKE_ESCAPE)
for q in qnorms
- )
+ ),
+ Page.target_path.in_(list(word_hits)),
),
)
)
@@ -479,9 +524,9 @@ async def search_paths_single(ctx: Any, query: str, limit: int) -> list[dict]:
continue
if is_excluded(target_path, spec):
continue
- scored.append(
- (max(_path_score(target_path, q) for q in qnorms), (page_id, title, target_path))
- )
+ substring = max(_path_score(target_path, q) for q in qnorms)
+ score = max(substring, word_hits.get(target_path, substring))
+ scored.append((score, (page_id, title, target_path)))
scored.sort(key=lambda pair: (-pair[0], pair[1][2]))
return [
diff --git a/tests/unit/server/mcp/test_search.py b/tests/unit/server/mcp/test_search.py
index 52f0b8edd..de59a51c8 100644
--- a/tests/unit/server/mcp/test_search.py
+++ b/tests/unit/server/mcp/test_search.py
@@ -167,6 +167,78 @@ async def test_path_search_preserves_non_trailing_glob_behavior(setup_mcp, query
assert [item["file"] for item in result] == expected
+_WORD_PATHS = [
+ "src/agents/command-poll-backoff.ts",
+ "lib/poll_backoff.py",
+ "web/pollBackoff.js",
+ "Services/PollBackoff.cs",
+ "net/poll.backoff.go",
+ "poll/backoff/index.ts",
+ "src/agents/poll.ts",
+ "src/agents/backoff-retry.ts",
+]
+
+
+async def _path_search(query, paths=_WORD_PATHS, limit=10):
+ import types
+
+ import repowise.server.mcp_server as mcp_mod
+ from repowise.server.mcp_server.tool_search_symbols import search_paths_single
+
+ for path in paths:
+ await _seed_page(f"file_page:{path}", path)
+ ctx = types.SimpleNamespace(session_factory=mcp_mod._session_factory, path="/tmp/test-repo")
+ return [item["file"] for item in await search_paths_single(ctx, query, limit=limit)]
+
+
+async def test_path_words_match_every_naming_shape(setup_mcp):
+ """Spaced words match kebab, snake, camel, Pascal and dotted filenames,
+ filename hits above a path whose directories carry the words."""
+ files = await _path_search("poll backoff")
+ assert set(files[:5]) == {
+ "lib/poll_backoff.py",
+ "web/pollBackoff.js",
+ "net/poll.backoff.go",
+ "Services/PollBackoff.cs",
+ "src/agents/command-poll-backoff.ts",
+ }
+ assert files[5:] == ["poll/backoff/index.ts"]
+
+
+ ("query", "expected"),
+ [
+ ("poll backof", "lib/poll_backoff.py"), # the last word may be a prefix
+ ("Poll Backoff", "lib/poll_backoff.py"),
+ ("pollBackoff", "lib/poll_backoff.py"),
+ ],
+)
+async def test_path_words_prefix_case_and_compound_queries(setup_mcp, query, expected):
+ files = await _path_search(query)
+ assert expected in files
+ assert "src/agents/poll.ts" not in files
+
+
[email protected]("query", ["poll jitter", "pol backoff", "agents retry backoff poll"])
+async def test_path_words_all_must_match(setup_mcp, query):
+ assert await _path_search(query) == []
+
+
+async def test_substring_hits_rank_above_word_hits(setup_mcp):
+ files = await _path_search("pollBackoff")
+ assert files[:2] == ["web/pollBackoff.js", "Services/PollBackoff.cs"]
+ assert "lib/poll_backoff.py" in files[2:]
+
+
+async def test_exact_path_query_is_unchanged(setup_mcp):
+ assert await _path_search("src/agents/poll.ts") == ["src/agents/poll.ts"]
+
+
class TestDecisionDownweight:
"""Decision records must not crowd file pages out of the top ranks."""