The idea
Run a vector search and a keyword search for the same question, then merge the two result lists. Vector search brings paraphrases; keyword search brings exact terms. Together they find more relevant chunks than either alone.
Scores from the two methods are on different scales, so merge by rank instead. Reciprocal Rank Fusion (RRF) gives each chunk 1 / (k + rank) from each list (k is usually 60) and sums them. Chunks that rank well in both lists rise to the top.
Retrieve generously at this stage, for example the top 20–50 from each, then let a reranker choose the final few for the prompt.
It is the most reliable single upgrade to a basic RAG pipeline and is used in nearly every production system.
Picture it
Words to know
- RRFReciprocal Rank Fusion: merge ranked lists by summing 1/(k+rank).
Try it without code
Merge two lists by hand
- Write two ranked lists of 5 doc IDs with some overlap.
- Score each ID with 1/(60+rank) per list and sum.
- See which IDs win.
Use it in code
Reciprocal rank fusion over any number of ranked ID lists.
from collections import defaultdict
def rrf(*ranked_lists, k=60):
scores = defaultdict(float)
for lst in ranked_lists:
for rank, doc_id in enumerate(lst, start=1):
scores[doc_id] += 1 / (k + rank)
return sorted(scores, key=scores.get, reverse=True)
vector_hits = [12, 7, 33, 4, 9]
keyword_hits = [7, 51, 12, 2, 33]
print(rrf(vector_hits, keyword_hits)[:5]) # 7 and 12 rise to the topImplement hybrid search over your pgvector table and compare Recall@5 against vector-only on your test questions.
Common mistakes
- Adding raw scores from different methods.
- Retrieving only 5 from each list, leaving nothing to rerank.