Query Expansion — Module 14.1

How This Lesson Fits the Module & Volume

Users ask short, vague, or vocabulary-mismatched questions. Query expansion rewrites or augments the query before retrieval to improve recall—complements hybrid search and re-ranking. It sits on the query side of the upcoming RAG pipeline.

Learning Objectives

By the end of this lesson, students should be able to:

Define query expansion and list common techniques (synonyms, multi-query, HyDE).
Explain how expansion trades recall gains for latency and noise.
Implement a simple multi-query retrieve-and-fuse pattern.
Contrast rule-based expansion with LLM rewriting.
Guard against query drift that retrieves irrelevant topics.
Evaluate expansion with recall@k on hard / short queries.

Definition

Query expansion improves retrieval by transforming a user query into one or more enriched queries (added terms, paraphrases, or hypothetical documents) whose search results are unioned or fused.

Technique Map

Technique	Idea	Risk
Synonym / glossary	Add domain aliases	Static lists go stale
Multi-query	LLM paraphrases → multi search	Cost × N queries
HyDE	Embed a hypothetical answer	Hallucinated topic drift
PRF	Expand from top hits’ terms	Feedback loop on bad hits

Helps

Short queries
Vocab mismatch
Multi-intent asks

Hurts

Already precise IDs
Tight latency budgets
Noisy corpora

Ops

Cache rewrites
Cap N paraphrases
Fuse with RRF

Code: Multi-Query + RRF Fuse

from collections import defaultdict
from sentence_transformers import SentenceTransformer, util

model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
corpus = ["PTO carryover rules", "Parental leave policy", "Sick leave documentation"]
emb = model.encode(corpus, normalize_embeddings=True, convert_to_tensor=True)

def search(q: str, k: int = 3):
    qe = model.encode(q, normalize_embeddings=True, convert_to_tensor=True)
    hits = util.semantic_search(qe, emb, top_k=k)[0]
    return [corpus[h["corpus_id"]] for h in hits]

def rrf_fuse(lists, k=60):
    scores = defaultdict(float)
    for lst in lists:
        for r, doc in enumerate(lst, 1):
            scores[doc] += 1.0 / (k + r)
    return sorted(scores, key=lambda d: -scores[d])

user_q = "time off after having a baby"
# In production, an LLM generates paraphrases; here we hardcode expansions:
expansions = [user_q, "parental leave", "maternity paternity PTO"]
fused = rrf_fuse([search(q) for q in expansions])
print(fused)

Strengths

Boosts recall on hard queries
Cheap wins with glossaries
Composes with hybrid + re-rank

Tradeoffs

Extra latency / embed calls
Drift retrieves wrong topics
Harder debugging

Common Misconception

“More paraphrases always help.” Past a point you retrieve a kitchen-sink of near-topic junk, burn tokens, and confuse the generator. Cap expansions, fuse ranks, and measure recall vs precision on a held-out set.

Knowledge Check

Short Answer: What is query expansion? Answer: Enriching/rewriting the query to improve retrieval recall.
True/False: HyDE embeds a hypothetical answer document. Answer: True (typically).
Multiple Choice: Multi-query cost scales with: (a) number of rewrites searched, (b) CSS files, (c) DPI. Answer: (a).
Short Answer: Name one drift risk. Answer: Expanded terms pull unrelated topics into the candidate set.
True/False: Exact SKU lookups usually need heavy paraphrasing. Answer: False—often hurts.
Multiple Choice: Fuse multi-query hits with: (a) RRF, (b) random delete, (c) only PNG. Answer: (a).
Short Answer: What is PRF? Answer: Pseudo-relevance feedback—expand using terms from initial top hits.
True/False: Expansion replaces the need for hybrid search. Answer: False—they complement.
Multiple Choice: Next lecture: (a) RAG Pipeline, (b) Vol. 1 only, (c) printers. Answer: (a).
Short Answer: How do you know expansion helped? Answer: Higher recall@k (without collapsing precision) on eval queries.

Key Takeaways

Expand queries to fix vocabulary mismatch and short asks.
Cap N, fuse with RRF, watch for topic drift.
Eval on hard queries; skip expansion when exact match dominates.
Next: RAG Pipeline.

Trainer’s Guide

Lab: Collect 15 failing short queries; add glossary expansion; measure recall@5 delta.

Prompt: Would you expand before or after metadata ACL filters? Why?

Recap: Query expansion is a recall lever on the question side. Continue with RAG Pipeline.