Search indexes

A search index is where your data is organised so it can be searched fast. You usually have one index per "kind" of thing — your product catalog, your articles, your support tickets, your knowledge base. An index isn't user-facing on its own — you wrap it in a search experience (or feed it to a chat experience as a tool) to put it in front of users.

Where to find these screens

Sidebar → Capabilities → Search Indexes.

The list screen

Lands on a list of every index you've created. From here you can:

• Filter by type (Lexical, Semantic, Hybrid) and by status (Ready, Creating, Indexing, Error, Offline).
• Search by display name or technical name.
• Switch between List and Grid view.
• Click New Index to create one.
• Click any row to open the index detail.

The stats cards at the top show the totals — how many indexes, how many are ready, how many documents in total across all of them.

Status badges, in plain English

Status	What it means
Ready	Index is up and accepting queries. Everything's good.
Creating	The index structure is being created. Happens for a few seconds after you click Create.
Indexing	Documents are being added or re-processed. Search still works on what's already there.
Error	Something went wrong. Open the index — the detail page shows a recovery banner.
Offline	Index exists but the underlying search service can't be reached. Usually a temporary connection issue.

Per-row actions

Hover any row to get the three-dot menu:

• View details — open the detail page.
• Edit — go to the edit screen.
• Reindex — rebuild from your current field config (see Rebuilding an index).
• Delete — destructive; asks you to type the name to confirm.

The three types of search

When you create an index, you pick how it searches. This cannot be changed later — the type is baked in.

Type	What it does	When to use it
Lexical	Keyword matching with the usual full-text smarts (stemming, fuzzy matching, boosting).	Fast, predictable, "did the user's words appear" search. SKUs, exact lookups, glossaries.
Semantic	Vector embeddings — finds documents close in meaning to the query, even when the exact words don't match.	"Comfortable shoes for long flights" should match a product titled "lightweight memory-foam loafers."
Hybrid	Both. Runs lexical and semantic in parallel and blends the rankings.	Default for most product catalogs and content libraries. Catches exact matches and meaning-based ones.

Semantic and hybrid indexes need an AI provider configured — the provider generates the embeddings. The embedding model and its vector size are locked at creation; switching later requires deleting and rebuilding because different models produce different-sized vectors.

Creating an index

Click New Index. A 2- or 3-step wizard runs.

Step 1 — Basic info

Field	What it controls
Display name	Human-readable name shown across the UI.
Index name	Auto-generated technical name. Used internally and can't be changed after creation.
Description	Optional. Helps your future self remember what this index is for.
Search type	Lexical, Semantic, or Hybrid (see above). Locked after creation.
Search provider	Where the index physically lives. Usually Elasticsearch (bundled). Can be Azure AI Search if you've configured it. Locked after creation.

Step 2 — Search settings

Field	What it controls
Indexing strategy	On Upload (default — documents are searchable immediately), Scheduled (batched), or Manual (only when you explicitly trigger). For most use cases, On Upload is right.
Language	Drives stemming and stop-word handling. Pick the language your content is in. Switching later requires a rebuild.
Synonyms	See Synonyms and stop words.
Custom stop words	See Synonyms and stop words.
Number of replicas (Elasticsearch)	How many copies of the index exist for redundancy. 0 is fine for development. Bumping it later doesn't need a rebuild.
Refresh interval (Elasticsearch)	How long between a document being uploaded and it being searchable. Default 1 second is fine.

Step 3 — AI configuration (Semantic and Hybrid only)

Field	What it controls
AI provider	Which provider generates the embeddings (e.g. OpenAI, Ollama).
Embedding model	The specific model. Different models produce different vector sizes. Locked after creation.
Vector similarity metric	How "close" two vectors are measured. Cosine is the right default — works with any normalised embedding model.

For Hybrid only, there are two extra settings:

Field	What it controls
RRF rank constant	How the rankings from lexical and semantic searches get merged. The default 60 is the right starting point.
RRF window size	How many top results from each search are considered for merging. Default 1000.

You usually don't need to touch these — the global defaults under Settings → Search Settings are sensible.

Click Create Index. You land on the detail page with an empty index. Next stop: Index fields.

The detail page

This is the home page for one index. Header gives you the name, the status badge, and primary actions:

• Upload — go to the document upload screen.
• Activate / Deactivate — toggle whether the index can serve queries.
• Export — download the configuration as JSON.
• Edit — open the edit screen.

Stats strip

Four numbers at the top:

• Documents — how many records are in the index right now.
• Index size — how much disk it's using.
• Fields — how many fields are mapped.
• Search type — Lexical / Semantic / Hybrid.

Index Fields card

Summary of the field mappings, with a link to the full Fields screen.

Index Name card

Technical name (copy button), plus a Reindex button — see Rebuilding.

Index information card

Template used, created date, updated date, last indexed timestamp, internal ID.

Search settings card

Read-only display of what you configured: indexing strategy, language, synonyms (if any), stop words (if any), provider-specific settings. Edit them on the Edit screen.

AI configuration card (Semantic / Hybrid only)

The AI provider, the embedding model, and the similarity metric. Has a warning that changing AI config requires re-uploading documents.

Error recovery banner

Only shown if the status is Error. Has a "Recreate Index" button that rebuilds the empty index structure from your saved field definitions, so you can re-upload your documents. Use this if the underlying search service got into a bad state.

Danger zone

Big red Delete button at the bottom. Requires you to type the index name to confirm.

The edit screen

Four tabs:

• General — display name, description, indexing strategy. The technical name, search type, provider, and template are shown but locked.
• Text Analysis — language, synonyms, stop words. Changing any of these shows a "Reindexing required" banner because the changes don't take effect for existing documents until you rebuild.
• Provider settings (Elasticsearch / Azure / …) — provider-specific knobs like replicas and refresh interval.
• AI Config (Semantic / Hybrid only) — read-only view of the current provider/model, with a Change Configuration button if you really need to switch. Switching deletes all indexed documents and requires you to re-upload — the dialog spells this out.

What a document looks like

A document is just a JSON object. You don't have to define a schema in advance — Interakt auto-detects fields when you bulk-load.

{
  "name": "Pacific runner sneaker",
  "brand": "Coastal",
  "category": "shoes",
  "price": 89.99,
  "in_stock": true,
  "description": "Lightweight running shoe with mesh upper..."
}

How those fields behave (searchable, filterable, displayed in results) is controlled in the Fields screen — see Index fields.

Common gotchas

• You can't change the search type after creating an index. If you started lexical and want hybrid, delete and recreate (the embeddings have to be generated from scratch anyway).
• You can't change the embedding provider on an existing semantic/hybrid index. Different providers produce different vector dimensions (768 for nomic-embed-text, 1024 for mxbai-embed-large, 1536 for text-embedding-3-small, 3072 for text-embedding-3-large). Pick deliberately at creation.
• Field type detection isn't always right. SKUs and model numbers often get flagged as text when they should be keyword. Fix in the Fields screen.
• Big bulk loads can fail mid-way. Check the error in the Recent Uploads panel; fix the bad documents; retry.
• Changing language, synonyms, or stop words doesn't affect existing documents until you rebuild — see Rebuilding an index.

Where to go next

• Index fields — facetable, searchable, sortable, boost, transforms — what every field setting does.
• Synonyms and stop words — text-analysis tuning.
• Rebuilding an index — when you need to, when you don't.
• Loading data — JSONL, CSV, batch sizes.
• Search experiences — how to put the index in front of users.