The Retrieval Playbook — 6 RAG Search Strategies

The 6 Plays

Pick the Right Strategy for the Query

Each method is designed for a specific kind of query signal. Using the wrong method is like ordering the wrong diagnostic test — you'll get results, just not the right ones.

💬 Vocab Gap

Dense Semantic Search

🧠

Converts your query into a vector and finds the closest-matching content — even when the words are completely different. Bridges the gap between how patients talk and how doctors write.

✓ Win:"I feel dizzy when I stand up quickly" → finds orthostatic hypotension — even though none of those words appear in the query.

✗ Miss:Exact drug names. "Losartan 50mg" might return related blood pressure drugs instead of the specific one.

🔑 Exact Keyword

Keyword (Full-Text) Search

🔑

Uses BM25 scoring — rare exact terms score highest. When the user knows the specific name, drug code, or clinical term, this method finds it with surgical precision.

✓ Win:"Metformin HCl 500mg" or "ICD-10 E11.9" → hits the exact article. One match out of thousands — found because the words exist verbatim.

✗ Miss:"Blood sugar medicine" → returns nothing useful if the database only uses clinical terminology.

👁 Visual

Visual (Multimodal) Search

👁️

Uses an embedding model that maps text descriptions and medical images into the same space. Patients can describe what they see — and the system finds matching images without knowing the diagnosis.

✓ Win:"Circular rash with a clear centre and red outer ring" → finds Lyme disease images. Patient didn't know the diagnosis; described what they saw.

✗ Need:Requires images to be embedded at storage time with a multimodal model (e.g. Gemini Embedding 2).

🔀 Visual + Filter

Hybrid Search

🔀

Applies a keyword filter first to narrow the field, then re-ranks what remains by visual or semantic similarity. Designed for queries carrying two different kinds of signals at once.

✓ Win:"Skin rash with raised red patches in tropical climates" — filter on tropical, re-rank by visual similarity to the rash description. Both signals used.

✗ Miss:If the filter keyword is too broad, it doesn't narrow enough. Too narrow, it filters out the correct result.

🎯 Known + Detail

Scoped Dense Search

🎯

User already knows the topic — they just want a specific detail inside it. Scope the search to the right document first, then do semantic search within only those chunks.

✓ Win:"What are the renal side effects of metformin?" — filter to the Metformin article, search inside it for kidney-related sections. The detail is buried mid-document.

✗ Need:Requires article name stored as metadata so the filter can scope the search correctly.

🤖 Compound

Agentic Search

🤖

An LLM (like Claude) reads the query, decides which combination of search tools to call, runs them, merges the results, and generates the final answer. For queries too complex to hardcode.

✓ Win:"Chest pain after eating fatty foods, worse lying down, 45-year-old male" — visual + symptom pattern + demographic. Claude decomposes and calls multiple tools.

⚠ Note:More powerful but less predictable — adds a layer of LLM decision-making. Build evals first.

💡

The rule that governs all of this: How you store your data determines which plays are available. If you don't embed images at ingestion, visual search isn't possible later. If you don't chunk by article section, scoped search won't work. Storage decisions come first — retrieval decisions follow.

Interactive

Try It: Query Decoder

Six real healthcare queries — each carrying a different signal. Click any to see which method wins, why the others miss, and what the correct result looks like.

↑ Click a query to decode it

Before You Ship

Your Minimum Viable Eval

You don't need 1,000 test cases. Start with 5–6 queries that reflect the real signal types your users will send. That's enough to know if you're improving.

Write 5–6 test queries

Cover the signal types you expect — at least one vocab-gap query, one exact keyword, and one that would fool the obvious default method.

Label what "correct" looks like

For each query, write down the article or passage you expect to surface. Three levels: did the right article appear? Did the right section appear? Was the answer answerable?

Run after every change

Every time you swap a method, adjust chunking, or change an embedding model — run the same 5–6 queries. Are you rank 1 more often? That's your signal of progress.

Which method to test first — a quick decision guide

💬 Vocab Gap

Non-experts asking questions in plain language, querying expert-written content (clinical notes, research papers, policy docs)

Start with Dense

🔑 Exact Term

Drug names, ICD codes, clinical identifiers, product SKUs — anything with a unique string that must appear verbatim

Start with Keyword

👁 Visual

Dermatology, radiology, wound care, pathology — anywhere users describe what they see before they know what it is

Start with Visual

🎯 Scoped

Users who know the document but need a detail from inside it — dosage sections, contraindications, specific procedure steps

Start with Scoped

🤖 Compound

Symptom checkers, triage assistants, diagnostic support tools — where one query carries symptoms + demographics + context simultaneously

Start with Agentic

Public Sources & Further Reading

Pinecone Documentation

docs.pinecone.io — hybrid search, namespaces, sparse-dense retrieval

Pinecone Learn

pinecone.io/learn — vector search concepts and developer guides

RAG Survey — Gao et al. 2023

arXiv:2312.10997 — comprehensive survey of RAG techniques

My workflow: A GenAcademy session introduced me to Arjun's work on production retrieval and set me down this research path. From there I went deep into Pinecone's public documentation and developer blog, read the RAG survey papers, and built this playbook by working through each retrieval strategy against real query types. This is my personal synthesis. A live session has far more depth — for the full curriculum, reach out to GenAcademy directly.