Keyword vs. Semantic vs. Hybrid Search: What to Choose
An employee asks “Can I send it back?” while the manual says “Returns policy.” Exact keyword search may miss it, while semantic search may retrieve the wrong A-104 product version. Choose methods by question type rather than assuming newer AI is always better.

Keyword Search matches terms while Semantic Search connects meanings
Keyword Search matches query terms against document wording, including model names, product codes, and legal text. Semantic Search uses numerical embeddings to retrieve related meanings even when wording differs. Hybrid Search combines both signals before ranking results (Google Cloud: Hybrid Search)
Hypothetical example: a manual says product replacement while a user asks for a new faulty item. Semantic search may find relevant passages without exact wording. But for “Which models support spare part XZ-17?”, matching the characters XZ-17 must match exactly; guessing which model is similar is insufficient.
Choose search methods around exact data requirements and intended meaning
- Documents containing identifiers and numbers: Prioritize exact keyword retrieval and filters such as product model, document year, or requirement code
- The user uses colloquialisms or synonyms: Add semantic search to recognize that “request a refund” and “return a payment” may share context
- Documents include both types: Consider Hybrid Search and check actual results with a set of user questions.
Google Cloud describes combining dense semantic embeddings with sparse keyword embeddings and merging rankings using methods such as Reciprocal Rank Fusion (Vertex AI: Hybrid queries) However, better rankings in one sample do not guarantee results for every Thai document.
Hypothetical example: The code matches, but the conditions may be in the wrong version.
Suppose a repository contains four documents: A is a general returns policy from 2026; B is a 2025 manual for model AB-018; C is a 2026 manual for AB-018 stating “Returns accepted within 7 days, except custom-made products”; and D is a 2026 manual for AB-081. The user asks, “Can I get a refund for a custom-made AB-018?” Document C, including its exception, must be retained. The model name, year, and meaning of the question all matter together.
Keyword search for “return products” may miss “accept returns”; semantic search may retrieve model D's manual because of similar content. Hybrid combines signals but does not guarantee C ranks first. For exact codes and years, use metadata filters or exact fields separately from similarity scores. Keyword scoring does not automatically validate every code character. All four documents are fictional; these are planning possibilities, not actual system results.
| Question | Details to preserve | Getting started |
|---|---|---|
| Within how many days can you return the item? | Period and exceptions | Semantic search, followed by checking the original text |
| Code AB-018 is used with which model? | The code must all match. | Keyword with exact field/code filter |
| Model service fee AB-018 2026 (B.E. 2569) | Both code and year | Hybrid with year filter |
This table is a proposed design approach, not results from any tested system. Specify expected documents before building, then measure retrieval frequency.
Check retrieved passages and generated answers separately
Semantic search can return similar meanings under different conditions; keyword search may miss spelling variants or synonyms. Hybrid search adds potential processing cost and ranking complexity. Embedding systems must also consider language, model version, document permissions, and chunking.
If connected to an AI answer generator, check Documents found separately from Answers written by the AI because the problems lie at different layers. Read the basics at What Is RAG, and Why Can It Still Answer Incorrectly?
Test realistic questions before choosing retrieval methods
- Collect 15–20 realistic questions, including synonyms, spelling variants, codes, numbers, and questions the documents cannot answer.
- Identify which source documents each question should retrieve, and keep document versions fixed during comparison.
- Run keyword, semantic, and hybrid search on identical data and check whether expected documents reach the top 3 or 5 results
- Record errors separated into “not found”, “found wrong version” and “found correct but the AI gave the wrong answer” before deciding to adjust the system.
Choose retrieval methods based on user questions and the consequences of a wrong result. No method is best for every task.
Sources
- Google Cloud: About hybrid search
- Google Cloud: Query public index with hybrid queries
- Google: source document
This article provides general knowledge. It is not individual financial, investment, legal, or tax advice. Readers are urged to review the latest information and consider their own circumstances before making any decisions.