AI & Vector Search
Opensolr's AI features go beyond traditional keyword matching. With vector search, your search engine understands the meaning behind every query — not just the words. Combined with AI Hints and AI Reader, your users get answers, not just links.
{!hybrid} parser
Hybrid search on Opensolr runs through {!hybrid}, a custom Solr query parser that fuses keyword and vector relevance correctly — it is what makes natural-language and cross-lingual queries work. See exactly how it works, and try live examples, on the Hybrid Search page.
New to AI search? Read the Hybrid Search deep-dive, try the live interactive demo, or follow the step-by-step testing guide.
How Hybrid Search Works
When a user types a query, Opensolr sends it through two paths at the same time and merges the results. This gives you the precision of keyword search plus the intelligence of AI semantic search.
What Is Vector Search?
In Plain English
Normal search works like a dictionary: it looks up the exact words you typed. If you search for “car repair”, it only finds pages that literally contain the words “car” and “repair”.
Vector search understands meaning. It knows that “car repair” is similar to “auto mechanic”, “vehicle maintenance”, and “fixing my automobile” — even though those pages never use the word “car” or “repair”. This means users find what they need, even when they do not know the exact terminology.
How It Works Under the Hood
Every piece of text — your documents and the user's query — gets converted into a list of 1,024 numbers (called a “vector”). Texts with similar meanings have similar numbers, so finding related content is as simple as finding nearby vectors.
Your documents and your users' queries go through the exact same AI model. That is why a query like “affordable restaurants” can match a document about “budget-friendly dining” — the AI produces similar number patterns for similar meanings. For the full technical details on how to create these embeddings via the API, see Bulk Embed All Documents. The vector itself lives in the embeddings field of your index, a 1,024-dimension knn_vector described in the Vector Search Schema Reference.
Hybrid Search: Best of Both Worlds
Opensolr does not make you choose between keyword search and vector search. Hybrid search runs both at the same time and merges the results. Here is why that matters:
Keyword Search
Finds exact matches. Great for product codes, names, technical terms. If a user searches “SKU-12345”, keyword search finds it instantly.
Vector Search
Finds similar meanings. Great for natural language questions. If a user asks “how to fix a leaky faucet”, vector search finds plumbing guides even if they never use the word “leaky”.
Hybrid (Both)
Gets the right answer regardless of how the user phrases their query. Precise when they are specific, intelligent when they are vague. The best of both worlds.
AI Hints
When a user searches your index, an AI-generated answer can appear above the search results. This answer is sourced entirely from YOUR content — it does not make things up. Think of it as a smart preview that saves your users time.
How the answer is put together
Retrieval and generation are two separate steps, and the first one is ordinary hybrid search — the same ranking, the same tuning and the same filters that produced the results on screen. The top four documents are then assembled into the prompt:
- each document is fenced and numbered, with an explicit end marker, so the model can see where one article stops and the next begins rather than reading four of them as one wall of text;
- the fragments the query actually matched are placed first, before the longer excerpt, so the focused text arrives while the model still has full attention on that document;
- documents scoring below half of the best hit are dropped — a narrow question otherwise arrives with one good match and three unrelated articles, and the model hedges;
- the question comes last, after the documents. Moving it to the front measurably collapses answer quality.
The instruction then asks for all the relevant documents, not just the closest one, and to combine what each adds into a single answer — which is why a broad question gets one coherent reply drawn from several articles instead of a summary of whichever one ranked first. When genuinely nothing in the index bears on the question, it says so plainly rather than inventing a fit.
Nothing is added from the model's own knowledge, and the temperature is deliberately low. If the answer is not in your index, you will not get one.
AI Hints are generated from the actual content in your Opensolr index. The AI reads your documents and summarizes the answer. It cites which document the answer came from so users can verify and click through.
AI Reader
Every search result gets a “Read” button. When a user clicks it, the AI fetches the full page content and retells it in its own words, right in the search page — a readable briefing in Markdown, not a one-line summary, so the reader gets what the page actually says without navigating away. It uses only what the page states; nothing is added from the model's own knowledge.
One-Click Summaries
Click “Read” on any result. The AI reads the entire page and produces a 2–3 paragraph summary highlighting the key points.
Streaming Response
The summary streams in word by word, so users see content appearing instantly. No waiting for the full response before text shows up.
API Endpoints
Developers can use the Opensolr AI and vector APIs directly. All endpoints are available via HTTPS and return JSON. Authentication is done with your API key.
Every call on this page runs against it immediately, with no signup: email mcp@opensolr.com, API key 420b8b23e7b12dc8ab838932145a5065. The index mcp_demo_d1__dense is preloaded with 300 news articles, so embed_and_search returns real hits on the first try.
Anything you create there is deleted after 3 days, automatically. The account is shared publicly — other people can change or delete what you create, so never put anything real or client-owned in it. The limits are per index and small on purpose: 200 MB bandwidth and 50 MB disk. For a private index that persists, get your own key — the free plan is free forever, no card.
/api/embed
Send a single piece of text and get back its 1,024-dimension vector embedding. Use this to convert text into a vector before storing or searching. Full API docs.
/api/batch_embed
Send up to 50 texts at once and get all their vectors back in a single response. Much faster than calling /api/embed 50 times. Full API docs.
/api/embed_and_search
The all-in-one endpoint. Send a query, and Opensolr converts it to a vector, runs the platform's tuned hybrid search, and returns ranked results — all in one API call. Your index's saved Search Tuning applies automatically, and every knob (field weights, minimum match, search mode, vector pool, quality boost) can be overridden per request.
/api/ai_summary
The instruction is the prompt. Build the whole thing yourself — your question and the
content you retrieved — and send it in one field. What you write is exactly what the model reads: nothing
is reordered, relabelled, appended or shortened on the way. Retrieval is yours too: run
embed_and_search, take the top hits, paste them into the prompt. Streaming or single-shot.
/api/image_to_text
Search by image. Upload a photo and get back the words that describe it — or, when the picture is a label, a receipt or a document, the text read off it — ready to use as an ordinary query. Nothing changes in your index — no image vectors, no new fields, no re-indexing — because what comes back is plain text that runs through the search you already have. Full API docs.
Code Example: embed_and_search
Here is a complete example that sends a natural language query and gets back semantically-ranked results from your index:
AI Enrichment Pipeline
When you index a document on a vector-enabled plan, Opensolr automatically enriches it with AI-generated metadata. This happens behind the scenes — you do not need to do anything.
A 1,024-dimension numerical representation of the document's meaning. This is what powers semantic similarity search.
Detects whether the document's tone is positive, negative, or neutral. Useful for filtering results by sentiment or sorting by tone.
Automatically identifies the language of each document (English, German, French, etc.). Enables language-aware search and filtering. See also: NLP Features & Named Entity Recognition.
AI and vector search features are available on plans that include vector indexing. If your current plan does not include vector search, contact support@opensolr.com to learn about upgrade options and pricing.
Related FAQ Articles
Hybrid Search Deep-Dive
Technical explanation of how keyword and vector search work together in Opensolr.
Live Demo: Try It Yourself
Interactive demo where you can test hybrid search against real data.
Single Embed API
API reference for generating a single vector embedding from text.
Batch Embed API
Embed up to 50 texts in a single API call for efficient bulk processing.
Related Documentation
Search & Embed
Learn about the hosted search page, embed code, and how to add search to your own website.
Search Tuning
Configure field boosts, relevance weights, and ranking strategies for optimal results.
API Reference
Full API documentation for all Opensolr endpoints including authentication and rate limits.