# Vector Search with Elasticsearch - Kibana Dev Tools Demo # # Environment: Elasticsearch Serverless, Vector Database project # Scenario: A catalog of Elastic products and solutions, each with # a short description and its official logo. # # Run the requests from top to bottom. Every step is safe to re-run. # Docs: https://www.elastic.co/docs/solutions/vector-database #### 0. Reset ############################################## # Start from a clean slate. On a fresh project these return # 404 "not found" - that's expected and harmless. force=true # drops the tool from any other agent that happens to reference # the same tool_id, so this stays idempotent even then. DELETE kbn:/api/agent_builder/agents/catalog-finder-agent DELETE kbn:/api/agent_builder/tools/catalog-search?force=true DELETE kbn:/api/agent_builder/tools/catalog-lookup-by-tag?force=true DELETE kbn:/api/agent_builder/tools/catalog-overview?force=true DELETE elastic-catalog?ignore_unavailable=true DELETE _ingest/pipeline/catalog-embed-description DELETE _inference/catalog-text-embeddings DELETE _inference/catalog-multimodal-embeddings #### 1. Inference API: Models on EIS ####################### # The Elastic Inference Service (EIS) comes with preconfigured, # ready-to-use endpoints. No model deployment, no API keys. GET _inference/_all?filter_path=endpoints.inference_id,endpoints.task_type # Create our own endpoints for two of the Jina models we'll use, # so we can give them memorable names. # Jina Embeddings v5 Text: text in, 1,024-dim vector out PUT _inference/text_embedding/catalog-text-embeddings { "service": "elastic", "service_settings": { "model_id": "jina-embeddings-v5-text-small" } } # Jina Embeddings v5 Omni: text and images in, mapped into the # same vector space. We'll use it on product logos. PUT _inference/embedding/catalog-multimodal-embeddings { "service": "elastic", "service_settings": { "model_id": "jina-embeddings-v5-omni-small" } } #### 1. Inference API: Invoking models ##################### # Text -> embedding (a list of 1,024 numbers) POST _inference/text_embedding/catalog-text-embeddings { "input": "a tool for building search experiences over your own content" } # Image -> embedding: a product logo POST _inference/embedding/catalog-multimodal-embeddings { "input": [ { "content": [ { "type": "image", "format": "base64", "value": "data:image/png;base64,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" } ] } ] } # What do vectors buy us? Similar meaning = high cosine # similarity, even with zero words in common. POST _query?format=txt { "query": """ ROW query = "a tool for building search experiences over your own content" | EVAL same_meaning = V_COSINE( TEXT_EMBEDDING("a tool for building search experiences over your own content", "catalog-text-embeddings"), TEXT_EMBEDDING("connectors and retrieval for custom search applications", "catalog-text-embeddings")), different_topic = V_COSINE( TEXT_EMBEDDING("a tool for building search experiences over your own content", "catalog-text-embeddings"), TEXT_EMBEDDING("detecting threats and hunting adversaries", "catalog-text-embeddings")) | KEEP query, same_meaning, different_topic """ } # Same idea across modalities: text vs. a product logo # (EMBEDDING is a preview function) POST _query?format=txt { "query": """ ROW logo = "Elastic Security logo" | EVAL matching_text = V_COSINE( EMBEDDING("a shield representing threat detection and SIEM", "catalog-multimodal-embeddings"), EMBEDDING("data:image/png;base64,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", "catalog-multimodal-embeddings", {"type": "image"})), unrelated_text = V_COSINE( EMBEDDING("a cloud representing managed hosting", "catalog-multimodal-embeddings"), EMBEDDING("data:image/png;base64,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", "catalog-multimodal-embeddings", {"type": "image"})) | KEEP logo, matching_text, unrelated_text """ } #### 2. Index: Embedding pipeline ########################## # "Bring your own vectors": an ingest pipeline that embeds each # product's description into a plain dense_vector field. PUT _ingest/pipeline/catalog-embed-description { "description": "Embed the product description with Jina v5 Text", "processors": [ { "inference": { "model_id": "catalog-text-embeddings", "input_output": [ { "input_field": "description", "output_field": "description_vector" } ] } } ] } #### 2. Index: Settings and mappings ####################### # Three ways to store vectors, side by side: # description_vector - dense_vector: you manage the embeddings # description_semantic - semantic_text: Elasticsearch chunks # and embeds text for you # logo - semantic: same, but for any modality PUT elastic-catalog { "settings": { "index": { "mode": "vectordb_document", "default_pipeline": "catalog-embed-description" } }, "mappings": { "properties": { "product_id": { "type": "keyword" }, "name": { "type": "text" }, "category": { "type": "keyword" }, "tags": { "type": "keyword" }, "logo_url": { "type": "keyword", "index": false }, "description": { "type": "text", "copy_to": "description_semantic" }, "description_vector": { "type": "dense_vector", "dims": 1024, "similarity": "cosine", "index_options": { "type": "bbq_disk", "auto_calibrate": true } }, "description_semantic": { "type": "semantic_text", "inference_id": "catalog-text-embeddings", "index_options": { "dense_vector": { "type": "bbq_disk" } } }, "logo": { "type": "semantic", "inference_id": "catalog-multimodal-embeddings" } } } } # Vector Database projects use the vectordb_document index # mode. It applies vector-tuned defaults, e.g. vectors are kept # out of _source to save storage. GET elastic-catalog/_settings?include_defaults=true&filter_path=**.mode,**.exclude_source_vectors # DiskBBQ keeps compressed vectors on disk, not in RAM. # auto_calibrate picks the cheapest encoding that reaches ~90% # recall for each merged segment of 10,000+ vectors (so it # won't kick in on this small demo). Note the bfloat16 default, # applied automatically by vectordb_document mode. GET elastic-catalog/_mapping?filter_path=**.description_vector #### 3. Load data ########################################## # Six Elastic products and solutions. Embeddings are generated # on the fly for all three vector fields - including each # product's logo. POST elastic-catalog/_bulk?refresh=wait_for&filter_path=errors,items.*.error {"index": {"_id": "elastic-elasticsearch"}} {"product_id": "elastic-elasticsearch", "name": "Elasticsearch", "category": "search", "tags": ["vector-search", "es-ql", "full-text"], "description": "A distributed search and analytics engine supporting full-text, vector, and hybrid search, with ES|QL for querying and dense_vector fields for semantic search.", "logo_url": "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAADAAAAAwCAIAAADYYG7QAAABC0lEQVR42u3Yv04CQRDH8Z1ZOGLAAkg8K/8lJNqdpTTnA2ilHT4H0lhZ8SCWdjwA25tQm1xiC8RIAiGKKzsW9DZEbkh+3yf4ZGeKyZKIyOJn8TzwLgvjmQliNhwT7+0W00bpNqGoQOHLzx96y9eRyTt7Gpcfr2zn7Nq7zChI3ufGMivRrPIu4zCe6QGF0ZRz2OK/RMJGWQABBNB/VyjdJKpAJCIYGUBrLTX1e3ghgAACCCCAthmEewigte+hrhvqWmpz/4KRAQQQQLmCLJMejWXio2qkB3RSi7iV1PSAWkmdO5f7FwdlDZrmYaWdxiQinz503fBp8PE2+V5u/GPfMh1Xo7vzejuNd4r8C9RvUSAo0hvEAAAAAElFTkSuQmCC", "logo": ["Elasticsearch", { "type": "image", "value": "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAADAAAAAwCAIAAADYYG7QAAABC0lEQVR42u3Yv04CQRDH8Z1ZOGLAAkg8K/8lJNqdpTTnA2ilHT4H0lhZ8SCWdjwA25tQm1xiC8RIAiGKKzsW9DZEbkh+3yf4ZGeKyZKIyOJn8TzwLgvjmQliNhwT7+0W00bpNqGoQOHLzx96y9eRyTt7Gpcfr2zn7Nq7zChI3ufGMivRrPIu4zCe6QGF0ZRz2OK/RMJGWQABBNB/VyjdJKpAJCIYGUBrLTX1e3ghgAACCCCAthmEewigte+hrhvqWmpz/4KRAQQQQLmCLJMejWXio2qkB3RSi7iV1PSAWkmdO5f7FwdlDZrmYaWdxiQinz503fBp8PE2+V5u/GPfMh1Xo7vzejuNd4r8C9RvUSAo0hvEAAAAAElFTkSuQmCC" }]} {"index": {"_id": "elastic-observability"}} {"product_id": "elastic-observability", "name": "Elastic Observability", "category": "observability", "tags": ["apm", "logs", "metrics", "uptime"], "description": "Unifies APM traces, infrastructure metrics, logs, and uptime checks in one place so teams can detect and resolve production issues faster.", "logo_url": "data:image/png;base64,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", "logo": ["Elastic Observability", { "type": "image", "value": "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAADAAAAAwCAIAAADYYG7QAAABbklEQVR42u3Xv0sCYRgH8O97d6bdmRZShwiFEAVWNESQa7QEUk6tLtbU2NBcU0tNrW1BS4PQH9CP2cEwMwQbKvB00OjqlLvzbWhKGnJIH+L9ji/PCx+el+eBl3HO8cdhu1kAssR0vxId8cYntNRCaFYf/Lm4Z6COrMeGj5PjkYCn41xCn5IpNGYO764fTSogAK9Nd/WkdFl+owIC8GG3N07LhmlTAQGovTvbmSdCIADn+XresAiBOMfRTZUQCMBFsfG1f6iAqqZzW7EIgQAUa01aoKppA1C6urO4tPybss10aiud6hZUt1xaHWpzTgtEa+wFSIAEqGdRADiFintfoQTKPbfOsuLJBEiABEiABEiA/i+ISRIxUMBHDBRSaYGUWBgSo9Qhv1ee1mlNmXdtjhbIE4/KU2OU9hBj6s4KUwcILUZJH9L2EyzoI7Sp5clR/0FSmY/0+V/2DRgOansJJ/diX5WcB4PXTN5yein6BPMxZ0y8HbS7AAAAAElFTkSuQmCC" }]} {"index": {"_id": "elastic-security"}} {"product_id": "elastic-security", "name": "Elastic Security", "category": "security", "tags": ["siem", "endpoint", "threat-hunting"], "description": "A SIEM and endpoint protection platform for detecting threats, hunting adversaries, and automating incident response across your environment.", "logo_url": "data:image/png;base64,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", "logo": ["Elastic Security", { "type": "image", "value": "data:image/png;base64,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" }]} {"index": {"_id": "elastic-search-apps"}} {"product_id": "elastic-search-apps", "name": "Elastic Search Applications", "category": "search-applications", "tags": ["connectors", "rag", "search-ui"], "description": "Tools for building search experiences and retrieval-augmented generation applications on top of your own content, with connectors to common data sources.", "logo_url": "data:image/png;base64,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", "logo": ["Elastic Search Applications", { "type": "image", "value": "data:image/png;base64,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" }]} {"index": {"_id": "elastic-cloud"}} {"product_id": "elastic-cloud", "name": "Elastic Cloud", "category": "deployment", "tags": ["serverless", "managed", "hosting"], "description": "Fully managed Elastic deployments, including serverless projects, so teams can run search, observability, and security workloads without operating infrastructure.", "logo_url": "data:image/png;base64,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", "logo": ["Elastic Cloud", { "type": "image", "value": "data:image/png;base64,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" }]} {"index": {"_id": "elastic-agent-builder"}} {"product_id": "elastic-agent-builder", "name": "Agent Builder", "category": "ai", "tags": ["agents", "es-ql-tools", "conversational"], "description": "Build conversational AI agents that answer questions over your own data by calling ES|QL tools, grounding every answer in real search results.", "logo_url": "data:image/png;base64,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", "logo": ["Agent Builder", { "type": "image", "value": "data:image/png;base64,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" }]} GET elastic-catalog/_count # Vectors are generated but not stored in _source GET elastic-catalog/_doc/elastic-security #### 4. Vector search with ES|QL ########################### # Baseline: keyword search. Incidental word overlap promotes # the wrong products over the one the user actually needs # (observability, for diagnosing a production incident). POST _query?format=txt { "query": """ FROM elastic-catalog METADATA _score | WHERE MATCH(description, "our production service is having problems and I need to find the cause") | SORT _score DESC | KEEP _score, product_id, name | LIMIT 5 """ } # kNN on dense_vector: embed the question, find the nearest # product vectors. observability jumps to the top. POST _query?format=txt { "query": """ FROM elastic-catalog METADATA _score | WHERE KNN(description_vector, TEXT_EMBEDDING("our production service is having problems and I need to find the cause", "catalog-text-embeddings")) | SORT _score DESC | KEEP _score, product_id, name | LIMIT 5 """ } # Filtered kNN: metadata filters are applied during the vector # search. visit_percentage is the DiskBBQ speed/recall knob. POST _query?format=txt { "query": """ FROM elastic-catalog METADATA _score | WHERE KNN(description_vector, TEXT_EMBEDDING("answer customer questions automatically using our own knowledge base", "catalog-text-embeddings"), {"visit_percentage": 10}) AND category in ("ai", "search-applications") | SORT _score DESC | KEEP _score, product_id, category, name | LIMIT 3 """ } # semantic_text: same result, much less work. No pipeline and # no query embedding - just MATCH the field. POST _query?format=txt { "query": """ FROM elastic-catalog METADATA _score | WHERE MATCH(description_semantic, "our production service is having problems and I need to find the cause") | SORT _score DESC | KEEP _score, product_id, name | LIMIT 5 """ } # The : operator is shorthand for MATCH. Combine with filters. POST _query?format=txt { "query": """ FROM elastic-catalog METADATA _score | WHERE description_semantic: "catch an attacker moving through our network" AND category == "security" | SORT _score DESC | KEEP _score, product_id, category, name | LIMIT 3 """ } #### 5. Multimodal search with the semantic field ########## # The semantic field is queried with Query DSL (ES|QL support # isn't available yet - ES|QL's KNN() explicitly rejects # "semantic" fields: "Cannot use field [logo] with unsupported # type [semantic]"). Because Agent Builder's ES|QL tools can # only run ES|QL, the catalog-finder-agent in section 7 cannot # answer "which logo looks like X" - that's only possible here, # directly against the index. # Text -> image: describe a shape, find the matching logo. POST elastic-catalog/_search { "_source": ["product_id", "name"], "size": 3, "query": { "match": { "logo": "a logo shaped like a shield" } } } # Image -> product: "which product has this logo?" - search # with a logo image instead of text. POST elastic-catalog/_search { "_source": ["product_id", "name"], "size": 3, "query": { "knn": { "field": "logo", "k": 3, "num_candidates": 20, "query_vector_builder": { "embedding": { "input": { "type": "image", "value": "data:image/png;base64,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" } } } } } } #### 6. Hybrid search with ES|QL ########################### # Exact terms like "SIEM" are where keyword search shines. # FORK runs keyword and semantic search in parallel; FUSE merges # them with Reciprocal Rank Fusion (RRF). POST _query?format=txt { "query": """ FROM elastic-catalog METADATA _id, _index, _score | FORK (WHERE MATCH(description, "SIEM threat detection") | SORT _score DESC | LIMIT 10) (WHERE MATCH(description_semantic, "SIEM threat detection") | SORT _score DESC | LIMIT 10) | KEEP _id, _index, _score, _fork, product_id, name | FUSE | SORT _score DESC | KEEP _score, product_id, name | LIMIT 5 """ } # Linear combination: normalize both score ranges to 0-1, then # weight them (here 30% keyword, 70% semantic). POST _query?format=txt { "query": """ FROM elastic-catalog METADATA _id, _index, _score | FORK (WHERE MATCH(description, "SIEM threat detection") | SORT _score DESC | LIMIT 10) (WHERE MATCH(description_semantic, "SIEM threat detection") | SORT _score DESC | LIMIT 10) | KEEP _id, _index, _score, _fork, product_id, name | FUSE LINEAR WITH { "weights": { "fork1": 0.3, "fork2": 0.7 }, "normalizer": "minmax" } | SORT _score DESC | KEEP _score, product_id, name | LIMIT 5 """ } #### 7. Agent Builder: ES|QL tools ######################### # Tool 1: hybrid search over the catalog POST kbn:/api/agent_builder/tools { "id": "catalog-search", "type": "esql", "description": "Search the Elastic product catalog with hybrid (keyword + semantic) search. Use for any question about what a product does or which product fits a use case.", "tags": ["catalog", "elastic"], "configuration": { "query": """ FROM elastic-catalog METADATA _id, _index, _score | FORK (WHERE MATCH(description, ?query) | SORT _score DESC | LIMIT 20) (WHERE MATCH(description_semantic, ?query) | SORT _score DESC | LIMIT 20) | KEEP _id, _index, _score, _fork, product_id, name, category, tags, description, logo_url | FUSE | SORT _score DESC | KEEP product_id, name, category, tags, description, logo_url | LIMIT 5 """, "params": { "query": { "type": "string", "description": "The user's need or question about Elastic products, in natural language" } } } } # Tool 2: exact lookup by tag or category POST kbn:/api/agent_builder/tools { "id": "catalog-lookup-by-tag", "type": "esql", "description": "Look up Elastic products by an exact tag or category, such as 'siem' or 'observability'.", "tags": ["catalog", "elastic"], "configuration": { "query": """ FROM elastic-catalog | WHERE MATCH(tags, ?tag) OR category == ?tag | KEEP product_id, name, category, tags, description, logo_url | LIMIT 10 """, "params": { "tag": { "type": "string", "description": "The exact tag or category to filter on, for example siem or observability" } } } } # Tool 3: what the catalog covers POST kbn:/api/agent_builder/tools { "id": "catalog-overview", "type": "esql", "description": "List the categories covered by the Elastic product catalog with product counts.", "tags": ["catalog", "elastic"], "configuration": { "query": """ FROM elastic-catalog | STATS products = COUNT(*), names = VALUES(name) BY category | SORT products DESC """, "params": {} } } # Try a tool directly, exactly as the agent will call it POST kbn:/api/agent_builder/tools/_execute { "tool_id": "catalog-search", "tool_params": { "query": "I want to let customers ask questions in natural language and get answers grounded in our data" } } #### 7. Agent Builder: Agent ############################### POST kbn:/api/agent_builder/agents { "id": "catalog-finder-agent", "name": "Elastic Product Finder", "description": "Recommends Elastic products and solutions from the catalog for a given use case.", "labels": ["catalog", "demo"], "avatar_color": "#BFDBFF", "avatar_symbol": "EF", "configuration": { "instructions": "You help people find the right Elastic product for a job, using only the product catalog. If the user names an exact tag or category, call catalog-lookup-by-tag first, then catalog-search for related options. Give a short recommendation with 1-2 sentences of reasoning and cite the product_id you used. When a product from the catalog is relevant to your answer, show its name and logo together as a Markdown image using its logo_url field, like ![Elastic Security](logo_url). Your tools can only match on product name, description, category and tags - they cannot search by what a logo looks like, so if asked to identify a product from a visual description of its logo, say that logo-shape search isn't something your tools support rather than guessing. If nothing in the catalog fits a request, say so plainly rather than guessing.", "tools": [ { "tool_ids": [ "catalog-search", "catalog-lookup-by-tag", "catalog-overview" ] } ] } } # Converse with the agent via API. The response shows which # tools it called and the ES|QL results it based its answer on. POST kbn:/api/agent_builder/converse { "agent_id": "catalog-finder-agent", "input": "I want to let customers ask questions in natural language and get answers grounded in our data - what do you recommend?" } # Keep chatting in the UI (replace with your # project's Kibana URL): # /app/agent_builder/agents/catalog-finder-agent/conversations/new # # Try asking: # - What's a good product for detecting threats in our network? # - I need to run my deployment without managing servers # - What categories of products do you have?