feat(embeddings): store knowledge bases at five vector widths and add Ollama

Knowledge bases were pinned to one 1536-dimension pgvector column, so only
models that emit exactly that width could index one. The embedding table now
carries a column per storable width — 384, 768, 1024, 1536, 3072 — and a base
records which one it uses at creation, chosen with EMBEDDING_OUTPUT_DIMS.

Ollama becomes an embedding provider on both paths: KB_EMBEDDING_MODEL accepts
ollama/<model> for knowledge bases, and the Embeddings block offers the models
installed on the configured server, with no API key and nothing billed.

Author: icecrasher321

apps/docs/content/docs/integrations/embeddings.mdx

@@ -13,19 +13,21 @@ import { BlockInfoCard } from "@/components/ui/block-info-card"
 {/* MANUAL-CONTENT-START:intro */}
 An embedding turns a piece of text into a list of numbers that captures its meaning. Two texts that mean similar things get similar numbers, so you can compare meaning directly instead of matching keywords. That is what powers semantic search, grouping related items, and spotting near-duplicates that are worded differently.
 
-The Embeddings block generates those numbers using OpenAI, Google Gemini, Cohere, or Mistral. Pick a provider, pick one of its models, pass in text, and get a vector back — one vector per input, in the order you supplied them. You can embed a single string or a list of strings in one call.
+The Embeddings block generates those numbers using OpenAI, Google Gemini, Cohere, Mistral, OpenRouter, or a model on your own Ollama server. Pick a provider, pick one of its models, pass in text, and get a vector back — one vector per input, in the order you supplied them. You can embed a single string or a list of strings in one call.
 
 Models differ in what they are good at and what they cost. `text-embedding-3-small` is the cost-efficient general choice, `gemini-embedding-001` gives the highest retrieval quality, `embed-v4.0` handles multilingual content, and `codestral-embed` is tuned for source code. Some models also let you trade vector size against quality, and some accept a task type so the vector is conditioned for how it will be used — the block only offers those controls for the models that actually support them.
 
 Two things worth knowing before you build on it. Vectors are only comparable when they come from the same model at the same size, so changing either means re-embedding everything you intend to compare. And input longer than the model's limit is shortened to fit rather than rejected, with a warning in the run, so chunk long documents yourself when the tail matters.
 
-Sim's knowledge bases embed separately, at a fixed vector width and from a smaller set of models. This block is for embedding text yourself inside a workflow.
+Ollama is the exception to most of the above. It runs on your own deployment, so it needs no API key and costs nothing, and the model list is whatever you have pulled onto that server rather than a catalog Sim maintains — the block reads it live, filters it to models that can actually embed, and shows each one's vector width next to its name. Ollama accepts neither a task type nor a size reduction, so the block does not offer those controls for it. Self-hosted deployments configure the server with `OLLAMA_URL`; on Sim Cloud there is no Ollama to reach, so the list comes back empty.
+
+Sim's knowledge bases embed separately: a base fixes one model and one vector width when it is created, from a smaller set of models. This block is for embedding text yourself inside a workflow.
 {/* MANUAL-CONTENT-END */}
 
 
 ## Usage Instructions
 
-Turn text into embedding vectors for semantic search, clustering, and similarity. Supports OpenAI, OpenRouter, Google Gemini, Cohere, and Mistral embedding models.
+Turn text into embedding vectors for semantic search, clustering, and similarity. Supports OpenAI, OpenRouter, Google Gemini, Cohere, and Mistral embedding models, plus any model on a self-hosted Ollama.
 
 
 
@@ -151,4 +153,25 @@ Generate embeddings from text using Mistral's embedding models
 | `dimensions` | number | Dimensionality of each vector |
 | `usage` | json | Token usage |
 
+### Ollama Embeddings
+
+Generate embeddings on a self-hosted Ollama server
+
+#### Input
+
+| Parameter | Type | Required | Description |
+| --------- | ---- | -------- | ----------- |
+| `input` | string | Yes | Text to embed, or an array of texts to embed in one call |
+| `model` | string | Yes | Embedding model pulled on the configured Ollama server |
+
+#### Output
+
+| Parameter | Type | Description |
+| --------- | ---- | ----------- |
+| `embeddings` | json | Generated embeddings |
+| `model` | string | Model used |
+| `provider` | string | Provider used |
+| `dimensions` | number | Dimensionality of each vector |
+| `usage` | json | Token usage |
+
 

apps/docs/content/docs/knowledgebase/index.mdx

@@ -23,6 +23,8 @@ When you upload a document, Sim processes it in the background:
 
 A document is searchable once its status reads `completed`. Open any document to view, edit, merge, or split its chunks.
 
+Every knowledge base records the embedding model and the vector width it was built with, and keeps them for its lifetime. Two bases built differently cannot be searched in one request — vectors are only comparable when they come from the same model at the same size — so search them separately, or rebuild one to match. On Sim Cloud every base uses `text-embedding-3-small` at 1,536 dimensions; self-hosted deployments choose both with [`KB_EMBEDDING_MODEL` and `EMBEDDING_OUTPUT_DIMS`](/platform/self-hosting/environment-variables), including models on their own Ollama.
+
 ## What you can upload
 
 Sim accepts PDF, Word, text, Markdown, HTML, Excel, PowerPoint, CSV, JSON, and YAML files, up to 100 MB each (best under 50 MB). Scanned PDFs work too: with Azure or [Mistral OCR](https://docs.mistral.ai/ocr/) configured, Sim extracts text from image-based pages.

apps/docs/content/docs/platform/self-hosting/environment-variables.mdx

@@ -55,7 +55,7 @@ import { Callout } from 'fumadocs-ui/components/callout'
 | `VERTEX_PROJECT` / `VERTEX_LOCATION` | Google Vertex AI — GCP project ID and region (region defaults to `us-central1`) |
 
 <Callout type="warn">
-  **Knowledge bases require a hosted embedding provider.** Three are supported, selected with `KB_EMBEDDING_MODEL`: `text-embedding-3-small` (default) and `text-embedding-3-large` on OpenAI or Azure OpenAI, and `gemini-embedding-001` on Gemini. There is no local embedding backend — configuring Ollama or vLLM does not substitute, because embeddings do not route through the configured chat model.
+  **Knowledge bases need an embedding model, selected with `KB_EMBEDDING_MODEL`.** Hosted: `text-embedding-3-small` (default) and `text-embedding-3-large` on OpenAI or Azure OpenAI, and `gemini-embedding-001` on Gemini. Local: any embedding model on your own Ollama, named `ollama/<model>` (for example `ollama/nomic-embed-text`) with `OLLAMA_URL` set. Ollama serves embeddings only through this variable — embeddings never route through the configured chat model, so vLLM and LM Studio do not substitute.
 </Callout>
 
 <Callout type="info">
@@ -190,10 +190,35 @@ See [Observability](/platform/self-hosting/observability).
 
 | Variable | Description |
 |----------|-------------|
-| `KB_EMBEDDING_MODEL` | Embedding model for new knowledge bases. Defaults to `text-embedding-3-small`; an unsupported value falls back to the default |
+| `KB_EMBEDDING_MODEL` | Embedding model for new knowledge bases. Defaults to `text-embedding-3-small`; use `ollama/<model>` for a model on your own Ollama. An unsupported value falls back to the default |
+| `EMBEDDING_OUTPUT_DIMS` | Vector width new knowledge bases are stored at: `384`, `768`, `1024`, `1536` (default), or `3072`. It must be a width the chosen model can emit; anything else falls back to `1536` with a warning |
 | `OPENROUTER_API_KEY` | Fallback route for the OpenAI embedding models — used when it is set and `OPENAI_API_KEY` is not the chosen path |
 | `COHERE_API_KEY` | Enables the Knowledge block reranker |
 
+Both variables apply at creation time and are recorded on the knowledge base, so changing either
+affects new knowledge bases only. Existing ones keep the model and width they were built with, and
+knowledge bases with different settings cannot be searched together.
+
+Matching the width to the model is yours to get right. Sim only knows the widths a hosted provider
+documents, and nothing at all about a model on your Ollama, so a mismatch surfaces as a failed
+document rather than a rejected configuration. Common pairings:
+
+| `EMBEDDING_OUTPUT_DIMS` | Works with |
+|----------|-------------|
+| `384` | `ollama/all-minilm` |
+| `768` | `ollama/nomic-embed-text`, `ollama/embeddinggemma`, `text-embedding-3-small`, `gemini-embedding-001` |
+| `1024` | `ollama/mxbai-embed-large`, `ollama/bge-m3`, `text-embedding-3-small`, `text-embedding-3-large` |
+| `1536` | `text-embedding-3-small`, `text-embedding-3-large`, `gemini-embedding-001` |
+| `3072` | `text-embedding-3-large`, `gemini-embedding-001` |
+
+`sim-setup add knowledge-embeddings` walks through all of this — pick OpenAI, Azure OpenAI, Gemini,
+or Ollama and it writes the variables that family needs. `sim-setup status` then reports the one
+family your `KB_EMBEDDING_MODEL` actually selects, rather than every provider you happen to hold a
+key for.
+
+The Embeddings block reads the same `OLLAMA_URL`. It lists the embedding-capable models installed on
+that server with the width each one emits, so a workflow can embed locally without an API key.
+
 ## Chat & PII
 
 | Variable | Description |

apps/docs/content/docs/platform/self-hosting/index.mdx

@@ -109,7 +109,7 @@ Sim is self-contained for the core editor and execution engine. A few features r
 
 | Feature | Requires | Notes |
 |---|---|---|
-| **Knowledge bases** | An OpenAI, Azure OpenAI, or Gemini API key | Embeddings are generated by a hosted provider, selected with `KB_EMBEDDING_MODEL` (`text-embedding-3-small` by default). There is no local embedding backend — knowledge bases are unavailable without one of these keys. |
+| **Knowledge bases** | An OpenAI, Azure OpenAI, or Gemini API key, or an Ollama server | Embeddings are generated by the model named in `KB_EMBEDDING_MODEL` (`text-embedding-3-small` by default). Set it to `ollama/<model>` with `OLLAMA_URL` to embed locally instead. `EMBEDDING_OUTPUT_DIMS` sets the vector width. |
 | **Agent blocks** | An API key for at least one model provider | Or a self-hosted OpenAI-compatible endpoint: Ollama, vLLM, LM Studio, or LiteLLM. |
 | **Chat module** | `COPILOT_API_KEY` from sim.ai | Set `NEXT_PUBLIC_CHAT_DISABLED=true` to hide the module instead. |
 | **Integrations** | Your own OAuth app per service | See [Integrations & OAuth](/platform/self-hosting/integrations-oauth). |

apps/docs/content/docs/platform/self-hosting/troubleshooting.mdx

@@ -250,7 +250,9 @@ The load balancer's backend timeout is closing them. On GKE, attach a `BackendCo
 
 ## Knowledge base upload fails
 
-Embeddings need a hosted provider — set `OPENAI_API_KEY`, configure Azure OpenAI, or set `KB_EMBEDDING_MODEL=gemini-embedding-001` with a Gemini key. There is no local embedding backend, so Ollama or vLLM does not substitute. If a key is set, verify pgvector is installed on the database.
+Embeddings need a provider — set `OPENAI_API_KEY`, configure Azure OpenAI, set `KB_EMBEDDING_MODEL=gemini-embedding-001` with a Gemini key, or set `KB_EMBEDDING_MODEL=ollama/<model>` with `OLLAMA_URL` to embed on your own Ollama. If one is configured, verify pgvector is installed on the database.
+
+A document that fails with `vector 0 has N unexpected dimensions` means `EMBEDDING_OUTPUT_DIMS` does not match what the model actually emits. Sim cannot check this for you — the message names both widths, so set the variable to the one the model returned and recreate the knowledge base. Existing knowledge bases keep the width they were created with.
 
 ## Credentials Unreadable After a Restore
 

apps/docs/content/docs/platform/self-hosting/verify.mdx

@@ -47,7 +47,7 @@ The full command list is in the [Docker guide](/platform/self-hosting/docker#the
 | 5 | Paste a model API key in settings and run a two-block workflow | Execution engine, credential encryption, outbound network | App logs; check `ENCRYPTION_KEY` is set and outbound egress is allowed |
 | 6 | Upload a small file in Files | File storage end to end | With object storage configured: presigned URL + bucket CORS. On local disk: the upload proxies through the app |
 | 7 | Upload a file larger than 50 MB | Multipart upload path (object storage only) | Check app logs for provider part-listing or completion errors |
-| 8 | Create a knowledge base and upload a PDF | Document parsing, embeddings, pgvector | Needs a hosted embedding provider — see below |
+| 8 | Create a knowledge base and upload a PDF | Document parsing, embeddings, pgvector | Needs an embedding provider — see below |
 | 9 | Invite a teammate from workspace settings | Email delivery | App logs for the mailer; see [Email](/platform/self-hosting/email) |
 | 10 | Connect an integration account | OAuth configuration | Redirect URI mismatch → see [Integrations & OAuth](/platform/self-hosting/integrations-oauth) |
 | 11 | Create a workflow with a Schedule trigger set to every minute, deploy it, wait 2 minutes | **Background jobs** | Check the scheduler's logs — see [Background Jobs](/platform/self-hosting/background-jobs) |
@@ -102,7 +102,7 @@ All six should be present on Compose: `simstudio`, `realtime`, `db`, `redis`, `c
 
 **Step 6 or 7 fails.** With object storage configured, a CORS error in the browser console means the bucket policy does not allow your Sim origin or the signed upload headers. If step 7 fails only during completion, check the app logs and verify the server identity can list multipart parts (for S3, `s3:ListMultipartUploadParts`). On local-disk storage there is no CORS involved — uploads proxy through the app, so look at the app logs and the proxy body-size limit instead.
 
-**Step 8 fails — knowledge base upload errors.** Knowledge bases need a hosted embedding provider — OpenAI, Azure OpenAI, or Gemini. There is no local embedding backend. If a key is set, check pgvector is installed on the database.
+**Step 8 fails — knowledge base upload errors.** Knowledge bases need an embedding provider: OpenAI, Azure OpenAI, or Gemini with an API key, or a model on your own Ollama via `KB_EMBEDDING_MODEL=ollama/<model>` and `OLLAMA_URL`. If one is configured, check pgvector is installed on the database.
 
 **Step 9 fails — no email arrives.** With no provider configured the mailer no-ops: it records the recipient, subject, and sender at `info` and reports success, never the message body. Raise `LOG_LEVEL` to `INFO` to see that line — the variable is uppercase-only, and at the production default of `ERROR` nothing is logged at all.
 

apps/sim/app/api/knowledge/search/utils.test.ts

@@ -48,7 +48,14 @@ afterEach(() => {
 })
 
 import { WORKSPACE_ACCESS_SCOPE } from '@/lib/knowledge/access/scope'
-import { generateSearchEmbedding } from '@/lib/knowledge/embeddings'
+import { generateSearchEmbedding, type KbEmbeddingTarget } from '@/lib/knowledge/embeddings'
+
+/** The platform default model and vector width, as a knowledge base records them. */
+const DEFAULT_EMBEDDING_TARGET: KbEmbeddingTarget = {
+  model: 'text-embedding-3-small',
+  dimensions: 1536,
+}
+
 import {
   executeKeywordSearch,
   executeKnowledgeSearch,
@@ -596,7 +603,7 @@ describe('Knowledge Search Utils', () => {
 
       mockNextEmbeddingResponse()
 
-      const result = await generateSearchEmbedding('test query')
+      const result = await generateSearchEmbedding('test query', DEFAULT_EMBEDDING_TARGET)
 
       expect(vi.mocked(fetch)).toHaveBeenCalledWith(
         'https://test.openai.azure.com/openai/deployments/text-embedding-ada-002/embeddings?api-version=2024-12-01-preview',
@@ -621,7 +628,7 @@ describe('Knowledge Search Utils', () => {
 
       mockNextEmbeddingResponse()
 
-      const result = await generateSearchEmbedding('test query')
+      const result = await generateSearchEmbedding('test query', DEFAULT_EMBEDDING_TARGET)
 
       expect(vi.mocked(fetch)).toHaveBeenCalledWith(
         'https://api.openai.com/v1/embeddings',
@@ -649,7 +656,7 @@ describe('Knowledge Search Utils', () => {
 
       mockNextEmbeddingResponse()
 
-      await generateSearchEmbedding('test query')
+      await generateSearchEmbedding('test query', DEFAULT_EMBEDDING_TARGET)
 
       expect(vi.mocked(fetch)).toHaveBeenCalledWith(
         'https://api.openai.com/v1/embeddings',
@@ -673,7 +680,7 @@ describe('Knowledge Search Utils', () => {
 
       mockNextEmbeddingResponse()
 
-      await generateSearchEmbedding('test query', 'text-embedding-3-small')
+      await generateSearchEmbedding('test query', DEFAULT_EMBEDDING_TARGET)
 
       expect(vi.mocked(fetch)).toHaveBeenCalledWith(
         'https://test.openai.azure.com/openai/deployments/custom-embedding-model/embeddings?api-version=2024-12-01-preview',
@@ -695,7 +702,7 @@ describe('Knowledge Search Utils', () => {
         OPENROUTER_API_KEY: undefined,
       })
 
-      await expect(generateSearchEmbedding('test query')).rejects.toThrow(
+      await expect(generateSearchEmbedding('test query', DEFAULT_EMBEDDING_TARGET)).rejects.toThrow(
         'OPENAI_API_KEY is not configured'
       )
     })
@@ -717,7 +724,9 @@ describe('Knowledge Search Utils', () => {
         text: 'Deployment not found',
       })
 
-      await expect(generateSearchEmbedding('test query')).rejects.toThrow('Embedding API failed')
+      await expect(generateSearchEmbedding('test query', DEFAULT_EMBEDDING_TARGET)).rejects.toThrow(
+        'Embedding API failed'
+      )
 
       // Clean up
       Object.keys(env).forEach((key) => delete (env as any)[key])
@@ -738,7 +747,9 @@ describe('Knowledge Search Utils', () => {
         text: 'Rate limit exceeded',
       })
 
-      await expect(generateSearchEmbedding('test query')).rejects.toThrow('Embedding API failed')
+      await expect(generateSearchEmbedding('test query', DEFAULT_EMBEDDING_TARGET)).rejects.toThrow(
+        'Embedding API failed'
+      )
 
       // Clean up
       Object.keys(env).forEach((key) => delete (env as any)[key])
@@ -756,7 +767,7 @@ describe('Knowledge Search Utils', () => {
 
       mockNextEmbeddingResponse()
 
-      await generateSearchEmbedding('test query')
+      await generateSearchEmbedding('test query', DEFAULT_EMBEDDING_TARGET)
 
       expect(vi.mocked(fetch)).toHaveBeenCalledWith(
         expect.any(String),
@@ -782,7 +793,7 @@ describe('Knowledge Search Utils', () => {
 
       mockNextEmbeddingResponse()
 
-      await generateSearchEmbedding('test query', 'text-embedding-3-small')
+      await generateSearchEmbedding('test query', DEFAULT_EMBEDDING_TARGET)
 
       expect(vi.mocked(fetch)).toHaveBeenCalledWith(
         expect.any(String),
@@ -811,7 +822,7 @@ describe('Knowledge Search Utils', () => {
       registry.recordResolved('TOKEN', 'secret-value')
 
       await runWithKnowledgeModelInputProvenance(registry, () =>
-        generateSearchEmbedding('prefix secret-value suffix', 'text-embedding-3-small')
+        generateSearchEmbedding('prefix secret-value suffix', DEFAULT_EMBEDDING_TARGET)
       )
 
       expect(vi.mocked(fetch)).toHaveBeenCalledWith(