apply Reranker class at chromaDB

llama_stack/providers/remote/vector_io/chroma/chroma.py

@@ -4,6 +4,7 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 import asyncio
+import heapq
 import json
 import logging
 from typing import Any
@@ -30,6 +31,7 @@ from llama_stack.providers.utils.memory.vector_store import (
     EmbeddingIndex,
     VectorDBWithIndex,
 )
+from llama_stack.providers.utils.vector_io.vector_utils import Reranker
 
 from .config import ChromaVectorIOConfig as RemoteChromaVectorIOConfig
 
@@ -161,7 +163,55 @@ class ChromaIndex(EmbeddingIndex):
         reranker_type: str,
         reranker_params: dict[str, Any] | None = None,
     ) -> QueryChunksResponse:
-        raise NotImplementedError("Hybrid search is not supported in Chroma")
+        """
+        Hybrid search combining vector similarity and keyword search using configurable reranking.
+        Args:
+            embedding: The query embedding vector
+            query_string: The text query for keyword search
+            k: Number of results to return
+            score_threshold: Minimum similarity score threshold
+            reranker_type: Type of reranker to use ("rrf" or "weighted")
+            reranker_params: Parameters for the reranker
+        Returns:
+            QueryChunksResponse with combined results
+        """
+        if reranker_params is None:
+            reranker_params = {}
+
+        # Get results from both search methods
+        vector_response = await self.query_vector(embedding, k, score_threshold)
+        keyword_response = await self.query_keyword(query_string, k, score_threshold)
+
+        # Convert responses to score dictionaries using chunk_id
+        vector_scores = {
+            chunk.chunk_id: score for chunk, score in zip(vector_response.chunks, vector_response.scores, strict=False)
+        }
+        keyword_scores = {
+            chunk.chunk_id: score
+            for chunk, score in zip(keyword_response.chunks, keyword_response.scores, strict=False)
+        }
+
+        # Combine scores using the reranking utility
+        combined_scores = Reranker.combine_search_results(vector_scores, keyword_scores, reranker_type, reranker_params)
+
+        # Efficient top-k selection because it only tracks the k best candidates it's seen so far
+        top_k_items = heapq.nlargest(k, combined_scores.items(), key=lambda x: x[1])
+
+        # Filter by score threshold
+        filtered_items = [(doc_id, score) for doc_id, score in top_k_items if score >= score_threshold]
+
+        # Create a map of chunk_id to chunk for both responses
+        chunk_map = {c.chunk_id: c for c in vector_response.chunks + keyword_response.chunks}
+
+        # Use the map to look up chunks by their IDs
+        chunks = []
+        scores = []
+        for doc_id, score in filtered_items:
+            if doc_id in chunk_map:
+                chunks.append(chunk_map[doc_id])
+                scores.append(score)
+
+        return QueryChunksResponse(chunks=chunks, scores=scores)
 
 
 class ChromaVectorIOAdapter(OpenAIVectorStoreMixin, VectorIO, VectorDBsProtocolPrivate):

llama_stack/providers/utils/vector_io/vector_utils.py

@@ -37,3 +37,114 @@ def sanitize_collection_name(name: str, weaviate_format=False) -> str:
     else:
         s = proper_case(re.sub(r"[^a-zA-Z0-9]", "", name))
     return s
+
+
+class Reranker:
+    @staticmethod
+    def _normalize_scores(scores: dict[str, float]) -> dict[str, float]:
+        """
+        Normalize scores to 0-1 range using min-max normalization.
+        Args:
+            scores: dictionary of scores with document IDs as keys and scores as values
+        Returns:
+            Normalized scores with document IDs as keys and normalized scores as values
+        """
+        if not scores:
+            return {}
+        min_score, max_score = min(scores.values()), max(scores.values())
+        score_range = max_score - min_score
+        if score_range > 0:
+            return {doc_id: (score - min_score) / score_range for doc_id, score in scores.items()}
+        return dict.fromkeys(scores, 1.0)
+
+    @staticmethod
+    def weighted_rerank(
+        vector_scores: dict[str, float],
+        keyword_scores: dict[str, float],
+        alpha: float = 0.5,
+    ) -> dict[str, float]:
+        """
+        Rerank via weighted average of scores.
+        Args:
+            vector_scores: scores from vector search
+            keyword_scores: scores from keyword search
+            alpha: weight factor between 0 and 1 (default: 0.5)
+                   0 = keyword only, 1 = vector only, 0.5 = equal weight
+        Returns:
+            All unique document IDs with weighted combined scores
+        """
+        all_ids = set(vector_scores.keys()) | set(keyword_scores.keys())
+        normalized_vector_scores = Reranker._normalize_scores(vector_scores)
+        normalized_keyword_scores = Reranker._normalize_scores(keyword_scores)
+
+        # Weighted formula: score = (1-alpha) * keyword_score + alpha * vector_score
+        # alpha=0 means keyword only, alpha=1 means vector only
+        return {
+            doc_id: ((1 - alpha) * normalized_keyword_scores.get(doc_id, 0.0))
+            + (alpha * normalized_vector_scores.get(doc_id, 0.0))
+            for doc_id in all_ids
+        }
+
+    @staticmethod
+    def rrf_rerank(
+        vector_scores: dict[str, float],
+        keyword_scores: dict[str, float],
+        impact_factor: float = 60.0,
+    ) -> dict[str, float]:
+        """
+        Rerank via Reciprocal Rank Fusion.
+        Args:
+            vector_scores: scores from vector search
+            keyword_scores: scores from keyword search
+            impact_factor: impact factor for RRF (default: 60.0)
+        Returns:
+            All unique document IDs with RRF combined scores
+        """
+
+        # Convert scores to ranks
+        vector_ranks = {
+            doc_id: i + 1
+            for i, (doc_id, _) in enumerate(sorted(vector_scores.items(), key=lambda x: x[1], reverse=True))
+        }
+        keyword_ranks = {
+            doc_id: i + 1
+            for i, (doc_id, _) in enumerate(sorted(keyword_scores.items(), key=lambda x: x[1], reverse=True))
+        }
+
+        all_ids = set(vector_scores.keys()) | set(keyword_scores.keys())
+        rrf_scores = {}
+        for doc_id in all_ids:
+            vector_rank = vector_ranks.get(doc_id, float("inf"))
+            keyword_rank = keyword_ranks.get(doc_id, float("inf"))
+
+            # RRF formula: score = 1/(k + r) where k is impact_factor (default: 60.0) and r is the rank
+            rrf_scores[doc_id] = (1.0 / (impact_factor + vector_rank)) + (1.0 / (impact_factor + keyword_rank))
+        return rrf_scores
+
+    @staticmethod
+    def combine_search_results(
+        vector_scores: dict[str, float],
+        keyword_scores: dict[str, float],
+        reranker_type: str = "rrf",
+        reranker_params: dict[str, float] | None = None,
+    ) -> dict[str, float]:
+        """
+        Combine vector and keyword search results using specified reranking strategy.
+        Args:
+            vector_scores: scores from vector search
+            keyword_scores: scores from keyword search
+            reranker_type: type of reranker to use (default: RERANKER_TYPE_RRF)
+            reranker_params: parameters for the reranker
+        Returns:
+            All unique document IDs with combined scores
+        """
+        if reranker_params is None:
+            reranker_params = {}
+
+        if reranker_type == "weighted":
+            alpha = reranker_params.get("alpha", 0.5)
+            return Reranker.weighted_rerank(vector_scores, keyword_scores, alpha)
+        else:
+            # Default to RRF for None, RRF, or any unknown types
+            impact_factor = reranker_params.get("impact_factor", 60.0)
+            return Reranker.rrf_rerank(vector_scores, keyword_scores, impact_factor)