ianarawjo/ChainForge v0.3.7

v0.3.7: RAGForge 🔎, In-browser LLMs 🌐, Doc Uploads (PDF, etc), OpenRouter 🔀, Reasoning Tokens, Grid View

on GitHub

ChainForge 0.3.7 is our biggest update in a while: RAG pipelines, LLMs that run in your browser with no API key, OpenRouter and reasoning models, and a refreshed Response Inspector.

🔎 RAGForge: retrieval-augmented generation

New Upload, Chunk, Retrieval and Rerank nodes let you build RAG pipelines and compare them the way you already compare prompts, e.g., two chunkers against three retrievers, side by side.

• Runs entirely in the browser in a "lite" model, with nothing installed: PDF, DOCX and Markdown uploads, BM25 and semantic search, and cross-encoder reranking.
• Adds ability to upload Documents using the Upload Docs Node, which can be used as input to prompt nodes, outside of RAG itself.
• Running locally with pip install chainforge[rag] adds more advanced RAG chunkers, retrievers, etc: server-side embeddings, LanceDB vector stores, TF-IDF, Chonkie chunkers and Cohere reranking. This is the recommended approach for developers---RAGForgeLite is just to let people get a taste on the web browser-only edition!

To get a sense of RAGForge, try the new Darwin's letters Q&A example (Example Flows → RAG Examples), which contains 60 public-domain letters from Charles Darwin.

💭 RAG Chat Node

Ask your pipeline questions in real time. The RAG Chat node runs each question through your flow's own nodes and shows one answer per configuration, labeled with what differs.

🌐 In-browser LLMs: no API key needed

Six small models now run entirely in your browser (via WebLLM): Gemma 3 1B, Llama 3.2 1B, Qwen2.5 0.5B, Qwen3.5 0.8B, SmolLM2 1.7B and Qwen3 1.7B. Each downloads once and is cached. Qwen2.5 0.5B is the new default model in Prompt Nodes. ChainForge also now supports Safari.

🔀 OpenRouter provider

• OpenRouter is a new provider: one API key for Claude, GPT, Gemini, Grok, DeepSeek, Qwen, Kimi and image models. Type in any model ID.

Reasoning tokens now appear!

• Reasoning is captured from Claude, Gemini, OpenAI, DeepSeek, OpenRouter and the in-browser Qwen models. It shows behind a 💭 toggle in the inspector, is available to evaluators as reasoning in the response.meta dict, and is sent back to the model in Chat Turns:

🔍 A refreshed Response Inspector

• Grid View lays responses out by any variables or by model, with filters, score badges and heatmap coloring. It stays fast with thousands of text or image responses.
• Table View and Grouped List use compact, model-colored response cards. Click one to read it in full.

🔗 See what's flowing along edges

Edges show how many values they carry. Hover over one to preview the first few values, their variables, and counts per model, without opening a node.

ALL Models and providers now UPDATED to 2026 editions!

• All models from OpenAI, Anthropic and Google, including GPT-5.x, Claude 5 and Gemini 3.x, and their image models (GPT Image 2.x, Gemini image models).
• HuggingFace now uses Inference Providers, since the old free Inference API no longer exists.
• Amazon Bedrock now uses the Converse API with a single settings form. Azure OpenAI uses Azure's v1 API. Together's model list is current, and MiniMax is a new provider.
• Aleph Alpha has been removed entirely, as it is no longer public.

On chainforge.ai/play

• API keys now persist upon reload of the page, using sessionStorage. They're kept for the tab, or remembered on your device if you opt in (localStorage).
• ChainForge asks your browser to keep your saved work, and warns you if the browser might clear it.
• Rename your flow from the top bar.

Running locally

• chainforge serve --idle-shutdown MINUTES stops the server when it hasn't been used for a while, and packager/ builds a Mac app that starts ChainForge on demand.
• Export flows as YAML for PromptStudio.

Thank you

Thanks to everyone who contributed to this release, including @dxa204 and @gauranshkumar (RAG nodes, Docker CI), @octo-patch (MiniMax), @micsco (Gemini SDK), @likeskymore (PromptStudio export), @ErnestHysa (bug fixes) and @cerebrixos (docs). 🎉