1. HyDE — Hypothetical Document Embeddings

Instead of embedding the short user query, generate a hypothetical answer and embed that. Bridges the gap between short queries and long documents.

def hyde_query(question, collection, llm):
    # Generate hypothetical answer
    hypo = llm.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "user",
                   "content": f"Write a detailed answer: {question}"}],
        max_tokens=200
    ).choices[0].message.content

    # Embed and search with the hypothetical doc
    return collection.query(query_texts=[hypo], n_results=5)

When to use: Vague queries, domain-specific corpus. Caution: adds latency, can hurt precision on specific/numerical queries.

2. Query Rewriting / Multi-Query

def multi_query(question, collection, llm, n=3):
    # Generate query variants
    variants = llm.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "user",
                   "content": f"Generate {n} different phrasings of: {question}"}]
    ).choices[0].message.content.strip().split('\n')

    # Search with each variant + original
    all_ids = set()
    all_docs = []
    for q in [question] + variants:
        r = collection.query(query_texts=[q], n_results=5)
        for id, doc in zip(r["ids"][0], r["documents"][0]):
            if id not in all_ids:
                all_ids.add(id)
                all_docs.append(doc)

    return all_docs[:10]  # Deduplicated top results

3. RAG Evaluation Metrics

Metric	What It Measures	How
Recall@K	Did the correct document appear in top K?	Ground truth annotation
MRR	How high did the correct doc rank?	1/rank of first correct result
Faithfulness	Is the answer grounded in retrieved context?	LLM-as-judge
Answer correctness	Is the final answer factually correct?	Human eval or reference answers

2026 Benchmark Results

Strategy	Recall@5	MRR@3
Hybrid + Reranker	0.816	0.605
Hybrid RRF	0.695	0.433
BM25 only	0.644	0.388
Dense only	0.587	0.355
HyDE	0.580	0.350

Achievements