From 0de1912117b69d83388c22cd46a6d19d2a5ce181 Mon Sep 17 00:00:00 2001 From: fzowl Date: Mon, 17 Aug 2026 01:46:38 +0200 Subject: [PATCH 1/2] Refresh VoyageAI embedding model catalog Add current VoyageAI models to both the Epsilla Cloud integration page and the vector-database embeddings reference, matching docs.voyageai.com: - voyage-4 family: voyage-4-large, voyage-4, voyage-4-lite, voyage-code-4 - contextualized: voyage-context-4 (voyage-context-3 already present / now added to the vector-database table) - multimodal: voyage-multimodal-3.5 Also fill gaps in the vector-database table (voyage-3.5, voyage-3.5-lite, voyage-multimodal-3) so it matches the integration page. Document that voyage-context-* inputs are embedded independently: each string is sent as its own document via a flat list with auto-chunking, yielding one deterministic vector per input; auto-chunking is disabled on the query path since the Voyage API rejects it for query inputs. --- platform/integrations/voyage-ai.md | 10 ++++++++++ vector-database/embeddings.md | 14 ++++++++++++++ 2 files changed, 24 insertions(+) diff --git a/platform/integrations/voyage-ai.md b/platform/integrations/voyage-ai.md index fcb8ab6..284a7c6 100644 --- a/platform/integrations/voyage-ai.md +++ b/platform/integrations/voyage-ai.md @@ -10,6 +10,12 @@ Epsilla integrates with Voyage AI with the following embedding models: | Name | Dimensions | |--------------------------------------|------------| +| **voyageai/voyage-4-large** | 1024 | +| **voyageai/voyage-4** | 1024 | +| **voyageai/voyage-4-lite** | 1024 | +| **voyageai/voyage-code-4** | 1024 | +| **voyageai/voyage-context-4** | 1024 | +| **voyageai/voyage-multimodal-3.5** | 1024 | | **voyageai/voyage-multimodal-3** | 1024 | | **voyageai/voyage-context-3** | 1024 | | **voyageai/voyage-3.5** | 1024 | @@ -26,6 +32,10 @@ Epsilla integrates with Voyage AI with the following embedding models: | **voyageai/voyage-large-2** | 1536 | | **voyageai/voyage-2** | 1024 | +{% hint style="info" %} +The `voyage-context-*` contextualized models are used the same way as any other embedding model. Each input string is embedded as its own independent document: Epsilla sends the batch as a flat list with auto-chunking enabled and a large chunk size, so every string resolves to exactly one chunk and one deterministic vector. Cross-input contextualization is not applied, because a batch may contain unrelated texts. On the query/retrieval path auto-chunking is disabled, since the Voyage API rejects it for query inputs. +{% endhint %} + For Epsilla open source vector db, you just need to add a header in the data ingestion and semantic search queries [like this](../../vector-database/embeddings.md#voyage-ai-embedding). Then you can start using the voyageai embedding models during vector table schema creation: diff --git a/vector-database/embeddings.md b/vector-database/embeddings.md index fcac3bb..6f960fc 100644 --- a/vector-database/embeddings.md +++ b/vector-database/embeddings.md @@ -292,6 +292,16 @@ Epsilla supports these VoyageAI embedding models (learn more about Voyage AI emb | Name | Dimensions | | ------------------------------------ | ---------- | +| **voyageai/voyage-4-large** | 1024 | +| **voyageai/voyage-4** | 1024 | +| **voyageai/voyage-4-lite** | 1024 | +| **voyageai/voyage-code-4** | 1024 | +| **voyageai/voyage-context-4** | 1024 | +| **voyageai/voyage-context-3** | 1024 | +| **voyageai/voyage-multimodal-3.5** | 1024 | +| **voyageai/voyage-multimodal-3** | 1024 | +| **voyageai/voyage-3.5** | 1024 | +| **voyageai/voyage-3.5-lite** | 512 | | **voyageai/voyage-large-2-instruct** | 1024 | | **voyageai/voyage-finance-2** | 1024 | | **voyageai/voyage-multilingual-2** | 1024 | @@ -310,6 +320,10 @@ Epsilla supports these VoyageAI embedding models (learn more about Voyage AI emb | **voyageai/voyage-3-lite** | 512 | | **voyageai/voyage-code-3** | 1024 | +{% hint style="info" %} +When you pick a `voyageai/voyage-context-*` contextualized model, each input string is embedded as its own independent document. Epsilla sends the batch as a flat list of strings with auto-chunking enabled and a large chunk size, so every string maps to exactly one chunk and one deterministic vector. Cross-input contextualization is intentionally not used, since a batch can contain unrelated texts. Auto-chunking is disabled on the query path because the Voyage API rejects it for query inputs. +{% endhint %} + When using Voyage AI embedding on Docker, make sure provide the **X-VoyageAI-API-Key** header when connecting to the vector database: {% tabs %} From 94e7e9093b7cdb81f0c9b0cda106b4a58cac44d7 Mon Sep 17 00:00:00 2001 From: fzowl Date: Mon, 17 Aug 2026 01:53:00 +0200 Subject: [PATCH 2/2] Address review: align voyage-context note, de-dupe model table - Rewrite the voyage-context-* hint to describe only user-observable behavior (each input embedded independently as its own document, one deterministic vector per input, no cross-batch contextualization) and drop unverified internal-processing details. Both docs pages now carry identical wording. - Remove the pre-existing duplicate voyage-code-2, voyage-law-2, voyage-finance-2, and voyage-multilingual-2 rows from vector-database/embeddings.md so the two model tables stay in sync. --- platform/integrations/voyage-ai.md | 2 +- vector-database/embeddings.md | 6 +----- 2 files changed, 2 insertions(+), 6 deletions(-) diff --git a/platform/integrations/voyage-ai.md b/platform/integrations/voyage-ai.md index 284a7c6..03033e9 100644 --- a/platform/integrations/voyage-ai.md +++ b/platform/integrations/voyage-ai.md @@ -33,7 +33,7 @@ Epsilla integrates with Voyage AI with the following embedding models: | **voyageai/voyage-2** | 1024 | {% hint style="info" %} -The `voyage-context-*` contextualized models are used the same way as any other embedding model. Each input string is embedded as its own independent document: Epsilla sends the batch as a flat list with auto-chunking enabled and a large chunk size, so every string resolves to exactly one chunk and one deterministic vector. Cross-input contextualization is not applied, because a batch may contain unrelated texts. On the query/retrieval path auto-chunking is disabled, since the Voyage API rejects it for query inputs. +The `voyageai/voyage-context-*` models are contextualized embedding models that you use exactly like any other embedding model in Epsilla. Each input string is embedded independently as its own document, so you get one deterministic vector per input. Contextualization across the other texts in the same batch is not applied, because a batch may contain unrelated documents. Pick these models for their standalone embedding quality. {% endhint %} For Epsilla open source vector db, you just need to add a header in the data ingestion and semantic search queries [like this](../../vector-database/embeddings.md#voyage-ai-embedding). diff --git a/vector-database/embeddings.md b/vector-database/embeddings.md index 6f960fc..10b870c 100644 --- a/vector-database/embeddings.md +++ b/vector-database/embeddings.md @@ -308,12 +308,8 @@ Epsilla supports these VoyageAI embedding models (learn more about Voyage AI emb | **voyageai/voyage-law-2** | 1024 | | **voyageai/voyage-code-2** | 1536 | | **voyageai/voyage-large-2** | 1536 | -| **voyageai/voyage-code-2** | 1536 | | **voyageai/voyage-2** | 1024 | | **voyageai/voyage-02** | 1024 | -| **voyageai/voyage-law-2** | 1024 | -| **voyageai/voyage-finance-2** | 1024 | -| **voyageai/voyage-multilingual-2** | 1024 | | **voyageai/voyage-lite-02-instruct** | 1024 | | **voyageai/voyage-3-large** | 1024 | | **voyageai/voyage-3** | 1024 | @@ -321,7 +317,7 @@ Epsilla supports these VoyageAI embedding models (learn more about Voyage AI emb | **voyageai/voyage-code-3** | 1024 | {% hint style="info" %} -When you pick a `voyageai/voyage-context-*` contextualized model, each input string is embedded as its own independent document. Epsilla sends the batch as a flat list of strings with auto-chunking enabled and a large chunk size, so every string maps to exactly one chunk and one deterministic vector. Cross-input contextualization is intentionally not used, since a batch can contain unrelated texts. Auto-chunking is disabled on the query path because the Voyage API rejects it for query inputs. +The `voyageai/voyage-context-*` models are contextualized embedding models that you use exactly like any other embedding model in Epsilla. Each input string is embedded independently as its own document, so you get one deterministic vector per input. Contextualization across the other texts in the same batch is not applied, because a batch may contain unrelated documents. Pick these models for their standalone embedding quality. {% endhint %} When using Voyage AI embedding on Docker, make sure provide the **X-VoyageAI-API-Key** header when connecting to the vector database: