know-cli 0.3.4__tar.gz → 0.3.6__tar.gz

{know_cli-0.3.4 → know_cli-0.3.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: know-cli
-Version: 0.3.4
+Version: 0.3.6
 Summary: Context Intelligence for AI Coding Agents — smart, token-budgeted code context
 Project-URL: Homepage, https://github.com/vic/know-cli
 Project-URL: Repository, https://github.com/vic/know-cli

{know_cli-0.3.4 → know_cli-0.3.6}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "know-cli"
-version = "0.3.4"
+version = "0.3.6"
 description = "Context Intelligence for AI Coding Agents — smart, token-budgeted code context"
 readme = "README.md"
 license = "MIT"

{know_cli-0.3.4 → know_cli-0.3.6}/src/know/lexical_index.py

@@ -217,8 +217,12 @@ class ChunkFTSIndex:
             cleaned = re.sub(r"[^A-Za-z0-9_]+", " ", q).strip()
             if not cleaned:
                 return q
-            # AND all terms for higher precision.
-            return " ".join(cleaned.split())
+            terms = cleaned.split()
+            if len(terms) == 1:
+                return terms[0]
+            # Use OR to avoid over-strict matching on multi-word queries.
+            # Ranking still handled by BM25.
+            return " OR ".join(terms)
 
         query_sanitized = _sanitize(query)
 

{know_cli-0.3.4 → know_cli-0.3.6}/src/know/semantic_search.py

@@ -684,7 +684,10 @@ class SemanticSearcher:
         # Hard fallback: plain text grep when hybrid/vector/lexical all miss.
         # This guarantees useful results for exact keyword queries in TS/React projects.
         if not results and query.strip():
+            import re
+
             q = query.strip().lower()
+            q_tokens = [t for t in re.findall(r"[a-z0-9_]+", q) if len(t) >= 3]
             exts = {".py", ".ts", ".tsx", ".js", ".jsx", ".go", ".rs", ".java", ".cpp", ".c", ".h", ".md"}
             fallback = []
             for p in root.rglob("*"):
@@ -703,16 +706,32 @@ class SemanticSearcher:
                     text = p.read_text(encoding="utf-8", errors="ignore")
                 except Exception:
                     continue
-                idx = text.lower().find(q)
+
+                text_l = text.lower()
+                # Prefer exact phrase, fallback to token hits.
+                idx = text_l.find(q)
+                if idx < 0 and q_tokens:
+                    token_positions = [text_l.find(tok) for tok in q_tokens]
+                    token_positions = [pos for pos in token_positions if pos >= 0]
+                    if not token_positions:
+                        continue
+                    idx = min(token_positions)
+
                 if idx < 0:
                     continue
+
                 line_no = text[:idx].count("\n") + 1
                 snippet = _format_snippet(p, line_no, line_no, context=2, include_line_numbers=include_line_numbers)
-                fallback.append((rel, line_no, snippet))
-                if len(fallback) >= top_k:
-                    break
+                # crude relevance: token coverage in file text
+                coverage = 0
+                if q_tokens:
+                    coverage = sum(1 for tok in q_tokens if tok in text_l)
+                fallback.append((coverage, rel, line_no, snippet))
+
+            fallback.sort(key=lambda x: (-x[0], x[1]))
+            fallback = fallback[:top_k]
 
-            for i, (rel, line_no, snippet) in enumerate(fallback, start=1):
+            for i, (coverage, rel, line_no, snippet) in enumerate(fallback, start=1):
                 results.append({
                     "schema": "know.search.v1",
                     "engine": "text-fallback",
@@ -723,7 +742,7 @@ class SemanticSearcher:
                     "chunk_type": "line",
                     "start_line": line_no,
                     "end_line": line_no,
-                    "scores": {"rrf": 0.01, "lexical_bm25": None, "vector": None},
+                    "scores": {"rrf": 0.01 + min(0.01, 0.002 * coverage), "lexical_bm25": None, "vector": None},
                     "snippet": snippet,
                     "rationale": {"lexical_available": lexical_available, "vector_available": vector_available},
                 })
@@ -834,8 +853,8 @@ class SemanticSearcher:
 
         results = []
         for i in top_indices:
-            if similarities[i] <= 0:
-                continue
+            # Keep top-k even when cosine is <= 0. In sparse/short queries,
+            # all scores can be near/under 0 and filtering causes empty results.
             chunk_key = file_paths[i]
             # Parse chunk key: "file_path::name::line_start"
             parts = chunk_key.split("::")