OpenAI components for building AI agents.
Changelog
19 versions
v2.0.0
Latest
- BREAKING: AI Agent: 'Model' and 'Instructions' moved from the component configuration to the 'in' port, so they can now be bound to flow variables (matching ai/claude and ai/gemini). Existing agents keep working: while the new inputs are left empty, the component keeps using the model and instructions from the old configuration. Fill in the new inputs to override them.
- SendPrompt: added an 'Instructions' input; the system message is no longer hardcoded to 'You are a helpful assistant.'.
- SendPrompt: added the 'Temperature', 'Max Tokens', 'Top P', 'Frequency Penalty' and 'Presence Penalty' inputs.
- GenerateEmbeddings, GenerateEmbeddingsFromFile: removed the 'embeddingTemplate' input, which was declared in the port schema but had no inspector field and was never read by the component.
- GenerateEmbeddings, GenerateEmbeddingsFromFile: the embeddings 'index' output is now correctly typed as an integer, and the required inputs declared in the schema now match what the components assert.
- TransformTextToJSON: fixed a crash that made the component fail on every execution ('Cannot access jsonSchema before initialization').
- TransformTextToJSON, VariableExtractor: the dynamic output-port schema request now passes ignoreAuth=true so the variable picker no longer needs an authenticated account.
- DescribeImages: the 'Model' input was nested inside the 'Images' field and therefore invisible in the Designer; it is now a top-level input.
- AI Agent: the 'Memory Data Store' input is now a typeahead so a store can be entered manually when the list cannot be loaded.
- Added output examples to every component with a static output port so the variable picker shows realistic sample values.
- CreateSpeech: fixed a crash that made the component fail on every execution ('context.httpRequest[method] is not a function'); the file extension now falls back to 'mp3' when no response format is set.
- Moderate: the 'Model' input is now honoured (it was declared but ignored) and the default model is 'omni-moderation-latest', as OpenAI no longer accepts 'text-moderation-latest'.
v1.4.1
- AI Agent: fixed OpenAI API error "Invalid value for 'content': expected a string, got null." that occurred when a tool chain produced no output. Tool results with null/undefined output are now coerced to a string before being sent to the model.
v1.4.0
- Added MakeApiCall component for performing arbitrary authorized API calls to the OpenAI API.
v1.3.4
- GenerateImage: handle b64_json response format for gpt-image-1/gpt-image-2 models; throw CancelError when API returns neither url nor b64_json.
v1.3.3
- Replaced direct Redis pub/sub with tenant-isolated context pubSubPublish method.
v1.3.2
- Fixed ioredis sentinel connection configuration to properly parse sentinel hosts and handle password authentication.
v1.3.1
- Fixed AI Agent mcpCallTool not passing correlationId, causing MCP server tool calls to appear in separate execution logs.
v1.3.0
- Introduced VariableExtractor component to specifically extract values for given variables from inputs using AI.
v1.2.3
- Updated langchain npm dependencies.
v1.2.2
- Added validation for required fields in components.
v1.2.1
- Renamed the AI Agent component label to remove the OpenAI suffix for clarity in the Designer.
v1.2.0
- Added support for files. AI agent now accepts a file that can be used as additional context for the prompt.
v1.1.2
- Fixed OpenAI API error "tool_call_ids did not have response messages" by ensuring all tool call IDs receive responses, even when tools timeout or fail. Added comprehensive error handling for malformed JSON arguments and improved polling logic to guarantee complete tool call coverage. Enhanced logging and reliability for AI conversations when tools experience issues.
v1.1.1
- Fixed a bug which prevented the AI Agent from using MCP server tools.
v1.1.0
- Add mcp output port for connecting MCP servers.
- Memory data store can be selected to store agent memory and summary.
- New connector configuration: AI_AGENT_MAX_HISTORY_SIZE (default 512000), llmBaseUrl and llmDefaultHeaders to support other OpenAI SDK compatible LLMs.
- Report usage and timing information in the output.
- Better performance, assistant API exchanged for chat completion.
- Support for streaming responses with built-in chat utility.
- Partial agent progress reported via streams too (for built-in chat utility).
v1.0.3
- Fixed an error `JSON schema is invalid` that occurred when the AiAgent used a tool with an empty parameters object.
v1.0.2
- TransformTextToJSON: improved error handling for invalid JSON schema input.
v1.0.1
- Fixed missing prompt in SendPrompt output variables. Other, small internal improvements.
v1.0.0
- First version.
Authentication
| Type | Details |
|---|---|
| API Key |
API Key (text)
Log into your OpenAI account and find API Keys page in the Settings.
|
Actions
13
Generate Embeddings From File
Generate embeddings for a text file. The text is split into parts, each part is split into chunks and embedding is returned for each chunk. The component emits embeddings array for each file part (1MB).
The returned embeddings is an array of the form:
TIP: use the JSONata modifier to convert the embeddings array into custom formats.
The returned embeddings is an array of the form:
[{ "index": 0, "text": "chunk1", "vector": [1.1, 1.2, 1.3] }].TIP: use the JSONata modifier to convert the embeddings array into custom formats.
Generate Embeddings
Generate embeddings for text data. The text is split into chunks and embedding is returned for each chunk.
The returned embeddings is an array of the form:
TIP: use the JSONata modifier to convert the embeddings array into custom formats. For convenience, the component also returns the first vector in the embeddings array which is useful when querying vector databases to find relevant chunks.
The returned embeddings is an array of the form:
[{ "index": 0, "text": "chunk1", "vector": [1.1, 1.2, 1.3] }].TIP: use the JSONata modifier to convert the embeddings array into custom formats. For convenience, the component also returns the first vector in the embeddings array which is useful when querying vector databases to find relevant chunks.