rivet

Bug Fixes

• Removed accidentally left in console.dir statements

New Features

• Added new Delegate Function Call node
• Assistant chat message type now can have function_calls for parallel function calling support.
• Add gpt-4o and gpt-4o-mini as possible models
• (Anthropic Plugin) Add claude-3-5-sonnet-20240620 as a possible model

Bug Fixes & Tweaks

• Correctly attach onGraphStart, onGraphError, and onGraphFinish for createProcessor. (#416)
• Chunk node correctly uses input port
• Slice node correctly uses input ports
• Updated AssemblyAI package (#427)

Bug Fixes

• Fix Vector Store and Vector KNN nodes not being able to use input ports for the collection ID.

New Features

• Support binary output in Http Call node
• Add internal support for latest gpt-4-turbo

New Features

• (Anthropic Plugin) Claude 3 support for tool usage.

Bug Fixes

None

New Features

• (Anthropic Plugin) Full Claude 3 support!
• Get Embedding node can be provided a model, and a number of dimensions to generate
• (AssemblyAI Plugin) Allow sending additional parameters via a JSON object
• (Pinecone Plugin) Allow object data to be passed and stored in the metadata

Fixes & Changes

• Update AssemblyAI library to latest
• Allow accessing project path via project metadata

Minor Changes

• Chat node can optionally not send presence_penalty and frequency_penalty if those parameters are unset
• Do not always override an assistant's tools list in Run Thread node
• Support sending a file name to OpenAI with the Upload File node
• Retry on OpenAI API 408 timeout
• Ignore empty event lines logging when reading event stream (less log spam)

Bug Fixes & Minor Changes

Not quite just a patch but should be backwards compatible, sorry!

• Fix regression in 1.14.1 causing hanging or rejected promises with object -> string conversions
• Add GPT-3.5 1106
• (Anthropic Plugin) Support Claude 2.1
• (Anthropic Plugin) Better error handling, won't retry client errors
• (OpenAI Plugin) Fix Run Thread Node not waiting for run to finish to completion
• Support GPT parallel function calling (Chat node will return function-calls instead of function-call when enabled)

Bug Fixes & Improvements

• Fix Destructure Node wrapping single-value results in arrays. If you do $.path and it only matches one thing, you will get "result" out instead of ["result"]
• Validate HTTP Call Node URLs before calling them
• Improve Object Node type inference - if the type being returned is an object, it will correctly return object[] instead of object to improve type coercion in later nodes.
• Improve type coercion for type: 'object' values. This is now treated more or less the same as an any (the actual type is inferred by the value's type). This improves coercion to string for arrays.
• Fix isSplitSequential being serialized to project file.

New Features

• Google Gemini Plugin.

Bug Fixes

• Packages updated.
• AssemblyAI plugin and package updated.
• Fix concurrent chat node with local LLM (#249)
• Fix creating objects with type and value properties.
• Fix abort/subscribe memory leak.

Bug Fixes

• Fix circular import with ESM importing (might be esbuild specific)

Fixes

• Fix base64 encoding when executing in Node.js, which fixes GPT-4 Vision being used there

Tweaks & Fixes

• Fixed NodeDatasetProvider not preserving options such as save when it does not find a data file.
• Added requireFile option to NodeDatasetProvider, so that it can error when it does not find a dataset file.
• Better coercion of messages input for Run Thread node.
• Fix a Run Thread node error message.
• Correct functions input data type for Run Thread node.
• Run Thread passes a message input for the "on message creation subgraph" like the docs say

Tweaks & Bug Fixes

• Fixed request token count being 0 when number of choices is unset (1)
• Account for GPT functions in token counting by approximating what OpenAI does to calculate their token counts.

New Features

• Support for new OpenAI GPT-4 Turbo model!
• Support for new OpenAI GPT-4 Vision model!
• Full support for new OpenAI Assistants API! Use the new OpenAI plugin.
• Switched from deprecated OpenAI function calling API to new OpenAI Tools API. Existing graphs may break because the new API requires a tool_id to be passed for function call responses.
• New Nodes:
  • Assemble Message - Used with GPT-4 Vision, this allows you to assemble a single chat message that contains both textual parts and image parts!
  • URL Reference - Used with GPT-4 Vision and the Assemble Message node, this allows you to reference a URL image instead of uploading an image.
  • Destructure - A simplified version of the Extract Object Path node, this allows you to extract multiple paths from an object at once, avoiding the need for multiple Extract Object Path nodes.
  • Replace Dataset - Replace the contents of a dataset with new values, or clear a dataset when no values are passed in.
• Support for new OpenAI seed parameter
• Support for forcing a Chat node to run a specific tool (OpenAI)
• Support for OpenAI JSON output mode
• If/Else and If node now outputs control-flow-excluded for unconnected ports by default, instead of undefined. Existing If/Else and If nodes still will have the old behavior.
• Read File node can now read files as binary, instead of only text.
• Remote debugger can pass context values in dynamicGraphRun

Changes & Fixes

• Fix loop controller infinite looping
• Loop controller node now allows non-default input ports to be "Not Ran". This will pass through to the output ports. If any "Default" input port is excluded, then the whole loop controller will not run.
• Fix ghost opened projects
• Code node now defaults to input1 and output instead of input and output
• Fixed node port types being incorrect
• Code now will error if it doesn't return the values it expects
• Fixed Extract Object Path erroring when the input value is undefined or null
• Arrays are now inferred correctly, such as with Trivet test inputs
• (internal) getTokenCountForMessages is now async
• ChatMessage is now a union of SystemChatMessage, UserChatMessage, AssistantChatMessage, and FunctionResponseChatMessage
• The message of a chat message is now a ChatMessagePart | ChatMessagePart[] to support images in chat messages.
• Text editors can now have placeholders and max length
• Added file path browser editor definition
• nodeExcluded event now includes inputs, outputs, and reason
• Loop controller node now consumes control-flow-excluded

New Features

• Release Rivet CLI - run rivet graphs from the command line easily!
• Node editors can mark a field as autoFocus
• Add ability for nodes to be enabled/disabled
• Add support for a "main graph" in a project, and made graphId optional when creating a GraphProcessor, which will run the main graph, if it's set for a project.
• Node ports can have descriptions
• Added chatNodeTimeout setting to Settings interface to set a timeout for the initial response from the LLM.
• Assemble Prompt node can now compute token count for all messages
• Added overrideMaxTokens to chat node, to override the max allowed tokens for a model
• Added "All Messages" output to Chat node
• Added optional input ports to GptFunctionNode to allow name, description, and schema to be passed in as inputs
• GptFunctionNode now supports {{interpolations}} like the Prompt node and Text nodes.
• Loop Controller node can now break at max iterations instead of always erroring
• Added Match node exclusive mode, only one matching regex will be passed through

Changes & Fixes

• Move most of createProcessor into @ironclad/rivet-core as createCoreProcessor. createProcessor in rivet-node still functions the same way.
• Moved Loop Controller max iteration tracking into the loop controller node itself.
• Prompt node will stringify the function call arguments if necessary
• AssemblyAI plugin now uses AssemblyAI SDK
• Added fake local-model model to models list

New Features

• Support legacy OpenAI models (0301)

Bug Fixes

• Fix sequential splitting not saving

Fixes & Tweaks

• Fix Split Sequentially not saving in graphs
• Support for legacy OpenAI models

New Features

• Add input ports to Trim Chat Messages Node for "max token count" & "from beginning"
• Add configurable rendering for User Input node - currently can choose between Preformatted and Markdown.
• Status bar now shows the execution time of your graph in addition to token count & cost.
• Loop controller now has nicer grouped port rendering!
• Add iteration output to loop controller (starts at 1)
• Add variable name output to Get Global and Set Global nodes

Changes & Bug Fixes

• Updated all dependencies to latest
• Fix some dropdowns occasionally not being visible
• Fix an empty header name causing chat node to fail
• Fixed scrolling test cases list
• Code node now gives better error when it doesn't output a valid object.
• HTTP Call node now gives better error when it's likely a CORS problem

New Features

• Add simplified streaming of graph events - an async generator via createProcessor(...).getEvents(spec) and an SSE implementation via createProcessor(...).getSSEStream(spec). See docs for more info.
• Add simplified streaming for a single node via createProcessor(...).streamNode(nodeIdOrTitle) (SSE)
• GraphProcessor now emits a finish event when it has either completed successfully or failed
• Add input ports to Trim Chat Messages Node for max token count & "from beginning"
• Add an iteration output to the loop controller node
• Add variable name output to get and set global nodes

Changes & Bug Fixes

• Updated all dependencies to their latest versions
• Fix empty header causing chat node to error
• Better code node error when it doesn't return something valid
• Better HTTP Call node errors