Merge pull request #6112 from MervinPraison/claude/issue-5051-20261028

fix: exclude compaction summary from session search scoring

src/praisonai-agents/praisonaiagents/session/sqlite_store.py

@@ -57,7 +57,8 @@ class SqliteSessionStore(DefaultSessionStore):
     # Bumped whenever ``_flatten`` changes what content lands in the index, so
     # a store built by an older release rebuilds its FTS rows once on upgrade.
     # v2: archived_messages included (Issue #5031).
-    INDEX_CONTENT_VERSION = 2
+    # v3: compaction summaries excluded (Issue #5051).
+    INDEX_CONTENT_VERSION = 3
 
     def __init__(
         self,
@@ -235,11 +236,16 @@ class SqliteSessionStore(DefaultSessionStore):
         Includes ``archived_messages`` (turns rolled out of the active window
         by ``retention="compact"``) ahead of the active ``messages`` so a
         compaction re-index preserves — rather than deletes — the raw turns,
-        keeping the whole conversation searchable (Issue #5031).
+        keeping the whole conversation searchable (Issue #5031). Compaction
+        summaries are skipped: search never scores them, so indexing them
+        would only spend the candidate allowance on misses (Issue #5051).
         """
         parts = []
         for msg in (*session.archived_messages, *session.messages):
             content = getattr(msg, "content", "")
+            metadata = getattr(msg, "metadata", None)
+            if isinstance(metadata, dict) and metadata.get("compaction"):
+                continue
             if content:
                 parts.append(str(content))
         return "\n".join(parts)
@@ -926,7 +932,7 @@ class SqliteSessionStore(DefaultSessionStore):
                 if not isinstance(msg, dict):
                     continue
                 content = str(msg.get("content", ""))
-                if not content:
+                if not content or self._is_compaction_summary(msg):
                     continue
                 lowered = content.lower()
                 score = 0.0

src/praisonai-agents/praisonaiagents/session/store.py

@@ -2401,6 +2401,17 @@ class DefaultSessionStore:
             snippet = snippet + "…"
         return snippet
 
+    @staticmethod
+    def _is_compaction_summary(msg: Dict[str, Any]) -> bool:
+        """True for the synthetic ``retention="compact"`` rollup message.
+
+        The summary paraphrases archived turns that are scanned verbatim, so
+        it is kept for context/replay but excluded from scoring — otherwise a
+        session's rank would depend on how often it was compacted (Issue #5051).
+        """
+        metadata = msg.get("metadata")
+        return isinstance(metadata, dict) and bool(metadata.get("compaction"))
+
     @staticmethod
     def _searchable_messages(data: Dict[str, Any]) -> List[Dict[str, Any]]:
         """Project a persisted session into the list of messages to scan.
@@ -2477,7 +2488,7 @@ class DefaultSessionStore:
                 if not isinstance(msg, dict):
                     continue
                 content = str(msg.get("content", ""))
-                if not content:
+                if not content or self._is_compaction_summary(msg):
                     continue
                 lowered = content.lower()
                 score = 0.0

src/praisonai-agents/tests/unit/session/test_sqlite_store.py

@@ -332,6 +332,53 @@ class TestArchivedRecall:
         hits = reopened.search("zx-9271-alpha")
         assert [h.session_id for h in hits] == ["old"]
 
+    @pytest.mark.parametrize("store_cls", [DefaultSessionStore, SqliteSessionStore])
+    def test_compaction_summary_not_scored(self, tmp_dir, store_cls):
+        # Issue #5051: the synthetic summary paraphrases archived turns that are
+        # already scored verbatim, so compaction count must not change rank.
+        from praisonaiagents.session.store import SessionData, SessionMessage
+
+        store = store_cls(session_dir=tmp_dir)
+        turn = ("user", "deploy token zx-9271-alpha")
+        self._seed_with_archive(store, "plain", archived=[turn], active=[])
+        compacted = SessionData(session_id="compacted")
+        compacted.archived_messages.append(SessionMessage(role=turn[0], content=turn[1]))
+        compacted.messages.append(
+            SessionMessage(
+                role="system",
+                content="[Session Summary]\nuser: deploy token zx-9271-alpha",
+                metadata={"compaction": True, "compacted_count": 1},
+            )
+        )
+        store._save_session(compacted)
+
+        hits = {h.session_id: h for h in store.search("zx-9271-alpha")}
+        assert hits["plain"].score == hits["compacted"].score
+        hit = hits["compacted"]
+        anchor = next(m for m in hit.messages if m["index"] == hit.anchor_index)
+        assert anchor["archived"]
+
+    def test_summary_only_matches_do_not_exhaust_candidates(self, tmp_dir):
+        # Unscored summaries must not be indexed, or they spend the bounded
+        # candidate allowance and hide a real match (Issue #5051, Greptile P1).
+        from praisonaiagents.session.store import SessionData, SessionMessage
+
+        store = SqliteSessionStore(session_dir=tmp_dir)
+        for i in range(6):
+            data = SessionData(session_id=f"compacted-{i}")
+            data.archived_messages.append(SessionMessage(role="user", content="hello"))
+            data.messages.append(
+                SessionMessage(
+                    role="system",
+                    content="[Session Summary]\n" + "pelican " * 20,
+                    metadata={"compaction": True},
+                )
+            )
+            store._save_session(data)
+        self._seed_with_archive(store, "real", archived=[], active=[("user", "a pelican flew by")])
+
+        assert [h.session_id for h in store.search("pelican", limit=1)] == ["real"]
+
 
 class TestIndexedSessionRoute:
     """Indexed gateway/agent routing lookups (Issue #2956)."""