LCORE-1723: Cross-Encoder Reranking for enhanced RAG

pyproject.toml

@@ -72,6 +72,8 @@ dependencies = [
     "jinja2>=3.1.0",
     # To be able to fix multiple CVEs, also LCORE-1117
     "requests>=2.33.0",
+    # Used for RAG chunk reranking (cross-encoder)
+    "sentence-transformers>=2.0.0",
     "datasets>=4.7.0",
     # Used for error tracking and monitoring
     "sentry-sdk[fastapi]>=2.58.0",

src/configuration.py

@@ -27,6 +27,7 @@
     OkpConfiguration,
     QuotaHandlersConfiguration,
     RagConfiguration,
+    RerankerConfiguration,
     RlsapiV1Configuration,
     ServiceConfiguration,
     SplunkConfiguration,
@@ -465,6 +466,13 @@ def okp(self) -> "OkpConfiguration":
             raise LogicError("logic error: configuration is not loaded")
         return self._configuration.okp
 
+    @property
+    def reranker(self) -> "RerankerConfiguration":
+        """Return reranker configuration."""
+        if self._configuration is None:
+            raise LogicError("logic error: configuration is not loaded")
+        return self._configuration.reranker
+
     @property
     def rag_id_mapping(self) -> dict[str, str]:
         """Return mapping from vector_db_id to rag_id from BYOK and OKP RAG config.

src/constants.py

@@ -192,6 +192,8 @@
 # Inline RAG constants
 BYOK_RAG_MAX_CHUNKS: Final[int] = 10  # retrieved from BYOK RAG
 OKP_RAG_MAX_CHUNKS: Final[int] = 5  # retrieved from OKP RAG
+# Score multiplier applied to BYOK chunks after cross-encoder reranking (Solr chunks unchanged)
+BYOK_RAG_RERANK_BOOST: Final[float] = 1.2
 
 # Solr OKP constants
 SOLR_VECTOR_SEARCH_DEFAULT_K: Final[int] = 5

src/models/config.py

@@ -1812,6 +1812,29 @@ class OkpConfiguration(ConfigurationBase):
     )
 
 
+class RerankerConfiguration(ConfigurationBase):
+    """Reranker configuration for RAG chunk reranking."""
+
+    enabled: bool = True
+    model: str = Field(
+        default="cross-encoder/ms-marco-MiniLM-L6-v2",
+        title="Reranker model",
+        description="Cross-encoder model name for reranking RAG chunks. "
+        "Defaults to 'cross-encoder/ms-marco-MiniLM-L6-v2' from sentence-transformers.",
+    )
+
+    # Private attribute to track if this was explicitly configured
+    _explicitly_configured: bool = PrivateAttr(default=False)
+
+    @model_validator(mode="after")
+    def mark_as_explicitly_configured(self) -> Self:
+        """Mark this configuration as explicitly set when instantiated from user input."""
+        if self.model_fields_set:
+            self._explicitly_configured = True
+
+        return self
+
+
 class AzureEntraIdConfiguration(ConfigurationBase):
     """Microsoft Entra ID authentication attributes for Azure."""
 
@@ -1970,6 +1993,12 @@ class Configuration(ConfigurationBase):
         "in rag.inline or rag.tool.",
     )
 
+    reranker: RerankerConfiguration = Field(
+        default_factory=RerankerConfiguration,
+        title="Reranker configuration",
+        description="Configuration for neural reranking of RAG chunks using cross-encoder.",
+    )
+
     @model_validator(mode="after")
     def validate_mcp_auth_headers(self) -> Self:
         """
@@ -2072,6 +2101,46 @@ def validate_rlsapi_v1_quota_configuration(self) -> Self:
 
         return self
 
+    @model_validator(mode="after")
+    def validate_reranker_auto_enable(self) -> Self:
+        """Automatically enable reranker when both BYOK and OKP RAG are configured.
+
+        When users have both BYOK (Bring Your Own Key) entries in byok_rag and OKP
+        (OpenShift Knowledge Platform) configured in the RAG strategies, automatically
+        enable the reranker if it's not explicitly disabled. This improves result
+        quality when multiple knowledge sources are available.
+
+        Returns:
+            Self: The validated configuration instance with reranker potentially enabled.
+        """
+        # Check if BYOK RAG entries are configured
+        has_byok = len(self.byok_rag) > 0
+
+        # Check if OKP is configured in either inline or tool RAG strategies
+        # pylint: disable=no-member
+        has_okp = (
+            constants.OKP_RAG_ID in self.rag.inline
+            or constants.OKP_RAG_ID in self.rag.tool
+        )
+
+        # If both BYOK and OKP are present and reranker is using default settings,
+        # ensure it's enabled for optimal results
+        if (
+            has_byok
+            and has_okp
+            and not self.reranker.enabled
+        ):
+
+            logger.info(
+                "Automatically enabling reranker: Both BYOK RAG (%d entries) and OKP "
+                "are configured. Reranking improves result quality when multiple "
+                "knowledge sources are available.",
+                len(self.byok_rag),
+            )
+            self.reranker.enabled = True
+
+        return self
+
     def dump(self, filename: str | Path = "configuration.json") -> None:
         """
         Write the current Configuration model to a JSON file.