> ## Documentation Index > Fetch the complete documentation index at: https://www.meilisearch.com/docs/llms.txt > Use this file to discover all available pages before exploring further. # Semantic vs hybrid search > When to use pure semantic search vs hybrid search, and how to tune the balance between keyword and vector results. Meilisearch supports three search modes controlled by the [`semanticRatio`](/capabilities/hybrid_search/advanced/custom_hybrid_ranking) parameter: pure keyword search, pure semantic search, and hybrid search. Each mode has strengths and weaknesses depending on your data and how your users search. This page helps you understand the tradeoffs and pick the right approach for your use case. ## The three search modes The `semanticRatio` parameter controls how Meilisearch blends keyword and semantic results: | Mode | `semanticRatio` | How it works | | ------------- | ------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Pure keyword | `0.0` | Meilisearch uses only [full-text](/capabilities/full_text_search/overview) matching. Results must contain the query terms (or close variants). No embedder is queried. | | Hybrid | `0.0 < ratio < 1.0` | Meilisearch runs both keyword and semantic search, then merges the results. Lower values favor keyword matches, higher values favor semantic matches. | | Pure semantic | `1.0` | Meilisearch uses only vector similarity. Results are ranked by how close their embeddings are to the query embedding. | ## When to use each mode ### Pure keyword search (semanticRatio = 0) Best when: * Users search for exact product names, SKUs, or identifiers * Your dataset contains structured data with specific terminology (legal documents, medical records) * You need deterministic, explainable results * You want to avoid the latency cost of generating query embeddings Example queries that work well with keyword search: * `"iPhone 15 Pro Max 256GB"` * `"error code 0x80070005"` * `"Moby Dick Herman Melville"` ### Pure semantic search (semanticRatio = 1) Best when: * Users describe what they need in natural language rather than using specific terms * Your content is homogeneous (all product descriptions, all articles, all Q\&A pairs) * Vocabulary mismatch is common (users say "laptop" but documents say "notebook computer") * You are building [conversational search](/capabilities/conversational_search/overview) or Q\&A features Example queries that work well with semantic search: * `"something to keep my coffee warm at my desk"` * `"how do I fix a leaky kitchen faucet"` * `"comfortable shoes for standing all day"` ### Hybrid search (0 \< semanticRatio \< 1) Best when: * Your users mix specific terms with natural language descriptions * Your dataset contains diverse content types * You want to catch both exact matches and conceptually relevant results * You are building ecommerce search, documentation search, or knowledge bases Example queries that benefit from hybrid search: * `"wireless ergonomic keyboard"` (keyword "wireless" + semantic "ergonomic") * `"python async database connection"` (technical terms + conceptual meaning) * `"red summer dress under $50"` (product attributes + style description) ## Tradeoffs ### Relevancy Hybrid search typically delivers the best overall relevancy for general-purpose applications. Pure keyword search excels for exact-match queries but misses conceptually similar results. Pure semantic search handles vocabulary mismatch well but may miss results that contain the exact query terms. ### Latency | Mode | Relative latency | Notes | | ------------- | ---------------- | ---------------------------------------------------------- | | Pure keyword | Fastest | No embedding generation needed | | Pure semantic | Moderate | Requires generating a query embedding | | Hybrid | Slowest | Runs both keyword and semantic search, then merges results | The latency difference depends on your embedder. Cloud-based embedders (OpenAI, Cohere) add network overhead for query embedding generation. Local embedders (HuggingFace) avoid network calls but use server CPU. ### Vocabulary mismatch handling This is where semantic search provides the most value. Consider a kitchenware dataset: | Query | Keyword results | Semantic results | | ------------------------------ | ------------------------------ | ------------------------------------------- | | `"spatula"` | Documents containing "spatula" | Documents about spatulas, turners, flippers | | `"something to flip pancakes"` | Few or no results | Spatulas, turners, griddle tools | | `"KitchenAid KFE5T"` | Exact product match | May return similar products instead | Hybrid search balances these scenarios. It returns the exact "KitchenAid KFE5T" match from keyword search while also surfacing conceptually relevant "pancake flipper" results from semantic search. ## Decision guide Use the following table to choose your starting `semanticRatio`: | Use case | Recommended ratio | Reasoning | | ---------------------------------- | ----------------- | --------------------------------------------------------- | | Ecommerce product search | `0.5` to `0.7` | Users mix product names with descriptive queries | | Documentation or knowledge base | `0.5` to `0.8` | Natural language questions benefit from semantic matching | | Code search | `0.0` to `0.3` | Exact token matching is critical for code | | Q\&A or support tickets | `0.7` to `1.0` | Users describe problems in varied language | | Catalog with SKUs and part numbers | `0.0` to `0.3` | Exact identifiers must match precisely | | Blog or article search | `0.5` to `0.7` | Mix of topic searches and specific queries | These are starting points. Test with real queries from your users and adjust based on the results you observe. ## When NOT to use hybrid search Hybrid search is not always the best choice. Consider pure semantic search (`semanticRatio: 1.0`) instead when: * **Image-only search**: if your data is purely visual (image catalogs with no text metadata), keyword search has nothing to match against. Use pure semantic search with [multimodal embeddings](/capabilities/hybrid_search/how_to/image_search_with_multimodal). * **Similarity-based use cases**: if you are building a recommendation system using the [`/similar` endpoint](/capabilities/personalization/getting_started/recommendations), you are already using pure vector similarity. Hybrid search does not apply. * **Pre-computed embeddings without text**: if you provide your own embeddings for non-textual content (audio, sensor data), there are no keywords to match. ## Next steps Fine-tune semanticRatio and test different configurations Learn more about Meilisearch's keyword search capabilities Overview of hybrid and semantic search in Meilisearch