@elizaos/plugin-ollama 2.0.0-beta.1 → 2.0.3-beta.2

package/README.md

@@ -1,17 +1,11 @@
 # Ollama Plugin
 
-This plugin connects [Ollama](https://ollama.com/) to ElizaOS so agents can use **local** LLMs for text, embeddings, and JSON object generation—without sending prompts to a third-party API.
-
-## Why this plugin exists
-
-- **Privacy and compliance:** All inference can stay on your LAN or laptop.
-- **Cost:** No per-token cloud billing for experimentation.
-- **Offline / air-gapped:** Works wherever Ollama runs, including containers and remote servers.
+This plugin connects [Ollama](https://ollama.com/) to elizaOS so agents can use **local** LLMs for text, embeddings, and structured output—without sending prompts to a third-party API.
 
 ## Requirements
 
 - [Ollama](https://ollama.com/) installed and reachable (same machine or network).
-- ElizaOS runtime with this package enabled.
+- elizaOS runtime with this package enabled.
 - At least one Eliza-1 model created, e.g. `ollama create eliza-1-2b -f packages/training/cloud/ollama/Modelfile.eliza-1-2b-q4_k_m`.
 
 ## Installation
@@ -26,7 +20,7 @@ Ensure Ollama is running:
 ollama serve
 ```
 
-Register the plugin on your character / app config (exact shape depends on your ElizaOS version):
+Register the plugin on your character / app config (exact shape depends on your elizaOS version):
 
 ```json
 "plugins": ["@elizaos/plugin-ollama"]
@@ -34,7 +28,7 @@ Register the plugin on your character / app config (exact shape depends on your
 
 ## Architecture: Vercel AI SDK + `ollama-ai-provider-v2`
 
-Handlers use **`generateText`** (completion, structured generation, and non-streaming fallbacks for schema-only stream contexts or toolChoice-only requests), **`streamText`** (plain SSE chat **or** `stream: true` **with native tools**), **`generateObject`**, and **`embed`** from the **`ai`** package, backed by **`ollama-ai-provider-v2`**.
+Handlers use **`generateText`** (completion, structured generation, and non-streaming fallbacks for schema-only stream contexts or toolChoice-only requests), **`streamText`** (plain SSE chat **or** `stream: true` **with native tools**), and **`embed`** from the **`ai`** package, backed by **`ollama-ai-provider-v2`**.
 
 ## Streaming chat (`TextStreamResult`)
 
@@ -45,7 +39,7 @@ This plugin returns **`TextStreamResult`** from **`streamText`** when **`stream:
 - **Plain chat:** no **`responseSchema`**, **tools**, or **`toolChoice`** — every text delta is yielded to **`textStream`** (normal SSE).
 - **Native tools:** **tools** are present — Ollama streams the chat request with tools on the wire. For **`RESPONSE_HANDLER`** / **`ACTION_PLANNER`**, **`useModel`**’s streaming path concatenates only **`textStream`** chunks into the string passed to **`parseMessageHandlerOutput`**, so this adapter **drains** model text deltas internally and **yields a single trailing chunk** of the first tool’s **`arguments`** JSON (the v5 plan). **Why:** prepending arbitrary streamed text would break **`JSON.parse`** on that accumulated string. Other **`TEXT_*`** types forward all text chunks and attach **`toolCalls`** on the result (parity with OpenAI/OpenRouter).
 
-**Still `generateText`:** **`stream: true`** with **`responseSchema`** only (no tools), e.g. nested **`FACT_EXTRACTOR`** during SSE — **`stream` may still be true** on params; we **do not throw** and we **log at debug** because structured **`format: json`** is not wired through **`streamText`** here.
+**Still `generateText`:** **`stream: true`** with **`responseSchema`** only (no tools), e.g. nested **`FACT_EXTRACTOR`** during SSE — **`stream` may still be true** on params; we **do not throw** and we **log at debug** because this adapter intentionally keeps structured **`format: json`** on the completion path.
 
 **Still `generateText`:** **`stream: true`** with **`toolChoice`** but **no** resolved **`ToolSet`** on the wire — **log at debug**, then **`generateText`**. **Why:** the AI SDK’s **`streamText`** path used here expects **`tools`** in the same request; `toolChoice` alone is invalid for streaming and is not produced by core v5 (Stage 1 always passes tools with `toolChoice`).
 
@@ -55,13 +49,13 @@ This plugin returns **`TextStreamResult`** from **`streamText`** when **`stream:
 |----------|-----------------|----------------------------------|------|-----|
 | `true` | yes | (any; schema dropped if both) | **`streamText`** + tools | Same surface as other provider plugins; Ollama v2 supports tools on streaming `/api/chat`. |
 | `true` | no | no schema, no `toolChoice` | **`streamText`** plain | **`TextStreamResult`** so `useModel` can forward SSE chunks. |
-| `true` | no | schema only | **`generateText`** | Structured **`format`** is not implemented on **`streamText`** in this adapter; nested schema calls must not throw. |
+| `true` | no | schema only | **`generateText`** | Structured **`format`** stays on the completion path in this adapter; nested schema calls must not throw. |
 | `true` | no | `toolChoice` only | **`generateText`** | **`streamText`+tools** requires a tool set; log explains misconfiguration. |
 | `false` / absent | (any) | (any) | **`generateText`** (or structured serialize) | Normal completion path; Stage 1 without inherited streaming uses this. |
 
 ### Why `ollama-ai-provider-v2` (not the old `ollama-ai-provider`)
 
-ElizaOS tracks **AI SDK 5/6**. Older `ollama-ai-provider` exposed **model specification v1**; current `ai` only accepts **v2+** models and throws:
+elizaOS tracks **AI SDK 5/6**. Older `ollama-ai-provider` exposed **model specification v1**; current `ai` only accepts **v2+** models and throws:
 
 `Unsupported model version v1 for provider "ollama.chat"`.
 
@@ -77,7 +71,7 @@ Environment variables (or character `settings` with the same keys—**why:** let
 |----------|---------|---------|
 | `OLLAMA_API_ENDPOINT` | `http://localhost:11434` (normalized to `…/api`) | Ollama HTTP API base. |
 | `OLLAMA_SMALL_MODEL` / `SMALL_MODEL` | `eliza-1-2b` | Small / fast text model. |
-| `OLLAMA_LARGE_MODEL` / `LARGE_MODEL` | `eliza-1-9b` | Larger text model. |
+| `OLLAMA_LARGE_MODEL` / `LARGE_MODEL` | `eliza-1-4b` | Larger text model. |
 | `OLLAMA_EMBEDDING_MODEL` | `eliza-1-2b` | Embedding model id. |
 | `OLLAMA_DISABLE_STRUCTURED_OUTPUT` | _unset_ | If `1` / `true` / `yes` / `on`, **disables** JSON-schema structured text (see below). |
 
@@ -88,7 +82,7 @@ Optional model overrides: `OLLAMA_NANO_MODEL`, `OLLAMA_MEDIUM_MODEL`, `OLLAMA_ME
 ```
 OLLAMA_API_ENDPOINT=http://localhost:11434/api
 OLLAMA_SMALL_MODEL=eliza-1-2b
-OLLAMA_LARGE_MODEL=eliza-1-9b
+OLLAMA_LARGE_MODEL=eliza-1-4b
 OLLAMA_EMBEDDING_MODEL=eliza-1-2b
 ```
 
@@ -99,7 +93,7 @@ OLLAMA_EMBEDDING_MODEL=eliza-1-2b
   "settings": {
     "OLLAMA_API_ENDPOINT": "http://localhost:11434/api",
     "OLLAMA_SMALL_MODEL": "eliza-1-2b",
-    "OLLAMA_LARGE_MODEL": "eliza-1-9b",
+    "OLLAMA_LARGE_MODEL": "eliza-1-4b",
     "OLLAMA_EMBEDDING_MODEL": "eliza-1-2b"
   }
 }
@@ -133,7 +127,6 @@ Set to `1`, `true`, `yes`, or `on` if:
 |------------|------|
 | `TEXT_*` | Chat / completion-style text. |
 | `TEXT_EMBEDDING` | Vector embeddings. |
-| `OBJECT_*` | `generateObject` with `output: "no-schema"` (best-effort JSON object). |
 | `RESPONSE_HANDLER` / `ACTION_PLANNER` | Same text path; v5 Stage 1 uses **messages + tools + `toolChoice`**; other pipelines may use **`responseSchema`** for JSON. |
 
 ### Text example
@@ -156,7 +149,7 @@ const embedding = await runtime.useModel(ModelType.TEXT_EMBEDDING, {
 
 ## Chat messages, native tools, and `toolChoice` (v5)
 
-ElizaOS v5 can call text models with **`messages`** (chat turns), **`tools`** (function / tool definitions), and **`toolChoice`** (for example `"required"` on the message-handler stage so the model must emit a planner tool call).
+elizaOS v5 can call text models with **`messages`** (chat turns), **`tools`** (function / tool definitions), and **`toolChoice`** (for example `"required"` on the message-handler stage so the model must emit a planner tool call).
 
 This plugin forwards those fields to the Vercel AI SDK **`generateText`** or **`streamText`** (when **`stream: true`** and tools are present) backed by **`ollama-ai-provider-v2`**, and when the call is “native shaped” (messages, tools, tool choice, or structured output), it may return a **`GenerateTextResult`-like object at runtime** or a **`TextStreamResult`** with **`toolCalls`** — TypeScript still types `useModel` text handlers as `string` for historical reasons.
 
@@ -186,16 +179,8 @@ This plugin forwards those fields to the Vercel AI SDK **`generateText`** or **`
 | **`streamText.textStream` failed** / **`AI_NoOutputGeneratedError`** / process exit after Ollama **500** | Ollama returned an error body (e.g. **insufficient system memory** for the model) during a **streaming** `/api/chat` call; AI SDK retried then failed. | Check logs for **`ollamaResponseBody`** (plugin now extracts it). Free RAM on the Ollama host, use a smaller model, or lower concurrency. Streaming errors occur while **`useModel`** consumes **`textStream`**, not only inside **`generateText`**. |
 | Connection errors | Ollama not running or wrong URL | `curl` the `/api/tags` endpoint; fix `OLLAMA_API_ENDPOINT`. |
 
-More context: [Ollama plugin docs](https://docs.eliza.ai/) (registry page: `plugin-registry/llm/ollama`).
-
-## API reference
-
-Ollama’s HTTP API: [Ollama API docs](https://github.com/ollama/ollama/blob/main/docs/api.md).
+More context: [elizaOS documentation](https://docs.elizaos.ai/).
 
 ## Changelog
 
 See **[CHANGELOG.md](./CHANGELOG.md)**.
-
-## License
-
-See [LICENSE](./LICENSE).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@elizaos/plugin-ollama",
-  "version": "2.0.0-beta.1",
+  "version": "2.0.3-beta.2",
   "type": "module",
   "main": "dist/node/index.node.js",
   "module": "dist/node/index.node.js",
@@ -28,6 +28,12 @@
         "default": "./dist/node/index.node.js"
       },
       "default": "./dist/node/index.node.js"
+    },
+    "./*.css": "./dist/*.css",
+    "./*": {
+      "types": "./dist/*.d.ts",
+      "import": "./dist/*.js",
+      "default": "./dist/*.js"
     }
   },
   "files": [
@@ -46,21 +52,20 @@
   },
   "sideEffects": false,
   "dependencies": {
-    "@ai-sdk/ui-utils": "^1.2.8",
     "ai": "^6.0.174",
-    "js-tiktoken": "^1.0.18",
-    "ollama-ai-provider-v2": "^3.5.0"
+    "ollama-ai-provider-v2": "^3.5.0",
+    "zod": "^4.4.3"
   },
   "devDependencies": {
     "@biomejs/biome": "^2.4.14",
-    "@elizaos/core": "2.0.0-beta.1",
+    "@elizaos/core": "2.0.3-beta.2",
     "@types/node": "^25.0.3",
     "bun-types": "^1.3.12",
     "typescript": "^6.0.3",
     "vitest": "^4.0.0"
   },
   "peerDependencies": {
-    "@elizaos/core": "2.0.0-beta.1"
+    "@elizaos/core": "2.0.3-beta.2"
   },
   "scripts": {
     "dev": "bun run build.ts --watch",
@@ -105,7 +110,7 @@
         "type": "string",
         "description": "Name or tag of the large-sized Ollama model to use for text and object generation.",
         "required": false,
-        "default": "eliza-1-9b",
+        "default": "eliza-1-4b",
         "sensitive": false
       },
       "OLLAMA_EMBEDDING_MODEL": {
@@ -132,12 +137,12 @@
         "type": "string",
         "description": "Fallback environment variable for specifying a large model if OLLAMA_LARGE_MODEL is not set.",
         "required": false,
-        "default": "eliza-1-9b",
+        "default": "eliza-1-4b",
         "sensitive": false
       }
     }
   },
-  "gitHead": "646c632924826e2b75c2304a75ee56959fe4a460",
+  "gitHead": "82fe0f44215954c2417328203f5bd6510985c1fc",
   "eliza": {
     "platforms": [
       "browser",