> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://developers.alephant.io/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://developers.alephant.io/_mcp/server.

# Index Settings

> Select the OpenSearch embedding model, plan a safe model swap, and configure retrieval-time indexing features.

Index Settings controls how Standard deployments encode and prepare documents for OpenSearch-backed retrieval. Only an Admin should change these deployment-wide settings, and every change should be tested against representative indexed content.

## Embedding model

OpenCore uses the selected embedding model to encode document chunks and queries. The page offers configured cloud providers, supported self-hosted models, and custom embedding-model definitions.

| Model property             | Why it matters                                                         |
| -------------------------- | ---------------------------------------------------------------------- |
| Model name                 | Selects the embedding implementation and tokenizer.                    |
| Dimension                  | Determines the vector shape of the OpenSearch index.                   |
| Normalize                  | Controls whether embeddings are normalized before indexing and search. |
| Query and passage prefixes | Applies model-specific prefixes to query and document text.            |

Index Settings requires Standard mode. Lite does not start OpenSearch, so it cannot run connector indexing or an embedding-model swap. In managed Cloud deployments, the embedding selector and related controls are read-only because the service manages them.

Configure a cloud embedding provider's credentials before selecting one of its models. Confirm model dimension, normalization, and prefixes when defining a custom model; they become part of the index contract.

## Model swap

Changing the embedding model or contextual-retrieval settings requires a full re-index. OpenCore reads existing chunks from the current OpenSearch index, re-embeds them with the new settings, and writes them to the target index.

| Strategy                               | Search during re-index                                                                | Switch behavior                                                                                                                               |
| -------------------------------------- | ------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------- |
| Re-index All Connectors Then Switch    | The current index continues serving queries                                           | Ports all indexable connector content, then promotes the new index                                                                            |
| Re-index Active Connectors Then Switch | The current index continues serving queries                                           | Limits the port to active connectors, then promotes the new index                                                                             |
| Switch Before Re-index                 | The future index serves queries immediately; search may be incomplete during backfill | Promotes the future index immediately without deleting the original OpenSearch corpus; a background port backfills it from the original index |

The first strategy is the default and safest choice. Re-embedding can take hours or days for a large corpus. During a normal or active-only swap, the existing model remains available until promotion completes.

The current implementation ports stored OpenSearch chunks into the new index and recalculates their embeddings; it does not refetch every source document from its connector. Forward indexing continues to protect newer writes while the backlog is copied.

Canceling an in-progress swap stops the port and abandons this rebuild target and its progress while keeping the previous embedding model current. The cancel path does not guarantee that a physical OpenSearch index is deleted. Do not start another swap while an immediate-switch backfill is still pending.

## Reranking

Reranking is separate from embedding and OpenSearch index construction. The current Index Settings page does not render a reranker selector, and its **Apply & Re-index** request does not send reranker fields.

Treat an embedding-model swap as a retrieval-candidate change, not as a reranker change. Verify answer quality after the new index is promoted, but do not expect this page to alter a deployment's separate retrieval-time reranking behavior.

## Contextual retrieval

Contextual Retrieval adds document-level context to each indexed chunk before embedding to improve hybrid-search relevance. Enabling it can substantially increase LLM and embedding cost.

The control is available only when the deployment has at least one language model. After enabling it, select a **Contextual Retrieval LLM**; OpenCore blocks **Apply & Re-index** when the toggle is on but no model is selected. Managed Cloud deployments keep this setting disabled.

Changing the contextual setting or its model uses the full re-index flow. When contextual enrichment changes, OpenCore rebuilds the chunk context and embeddings in the target OpenSearch index rather than copying the old vectors.

## Advanced settings

| Setting                     | Availability                                        | Effect                                                                                         |
| --------------------------- | --------------------------------------------------- | ---------------------------------------------------------------------------------------------- |
| Multipass Indexing          | Temporarily unavailable; the switch is disabled     | Intended to index varying chunk sizes, but cannot currently be enabled from this page          |
| Extract & Caption Images    | Requires a configured vision-capable language model | Extracts images from supported uploaded files and adds generated descriptions during indexing  |
| Captioning LLM              | Available only after image processing is enabled    | Selects the vision-capable model used for new image captions                                   |
| Max Image Size for Analysis | Available only after image processing is enabled    | Skips images above the selected 5, 10, 20, 50, or 100 MB limit; the displayed default is 20 MB |

Captioning-model and image-processing updates apply to documents indexed afterward. Existing captions are already part of prior embeddings, so changing the captioning model alone does not rewrite them.

## Verify

1. Apply the new settings and confirm that the page reports **Re-indexing in progress** and names the target model.
2. Monitor the connector/indexing status. For a normal or active-only swap, confirm that queries continue using the previous model until the new index is promoted.
3. After promotion, run representative keyword, semantic, and hybrid searches against content described in [Connectors & Indexing](/opencore/knowledge/connectors).
4. Confirm that the current embedding model has changed, expected documents are searchable, and contextual or image features behave only under the prerequisites described above.
5. If the deployment has no OpenSearch-backed indexing controls, verify its mode in [Deployment Modes](/opencore/deployment-modes) and configure a model through [Language Models](/opencore/models/language-models) before retrying dependent features.