AI / LLM and RAG nodes

Call large language models from a graph and build retrieval-augmented generation (RAG) pipelines — answering questions from your own documents. Source: backend/core/nodes/ai/ai_nodes.py. Demo: projects/demo_project/ai/rag_demo.weave.

Load Documents ─▶ Split Text ─▶ Build Vector Index ─▶ Retrieve ─▶ Format Context ─┐
                                                          ▲                       ▼
                                                    question ───────────▶ Prompt Template ─▶ LLM Chat

Nodes

Node	Category	What it does
LLM Chat	AI / LLM	Sends a prompt (+ optional system text) to OpenAI, Anthropic Claude, Google Gemini, Ollama or an OpenAI-compatible server; outputs the reply. Model blank = the provider's default.
Prompt Template	AI / LLM	Fills {question}, {context} and any {key} from a values dict. Unknown placeholders are left as written.
Load Documents	AI / RAG	Reads text files in a folder matching a glob (**/*.md, **/*.txt).
Split Text	AI / RAG	Chunks documents by paragraphs, markdown sections or fixed character windows, with overlap.
Embed Text	AI / RAG	Texts → an (n, d) array of unit vectors (for clustering / plotting with the ML nodes).
Build Vector Index	AI / RAG	Indexes chunks for similarity search; a plain dict (method, model, chunks, matrix, vectorizer).
Retrieve	AI / RAG	Top-k chunks for a query by cosine similarity.
Format Context	AI / RAG	Numbers and labels retrieved chunks ([1] source) into one context string, up to a character limit.

Providers and keys

LLM Chat uses the same providers and .env keys as the chat assistant: OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY, OLLAMA_HOST (optional) and OPENAI_COMPATIBLE_BASE_URL / _API_KEY / _MODEL. Keys are never stored in the graph, and exported scripts read them from the environment (or a .env found from the working directory). Every run of an LLM Chat node is a real, possibly billed, API call.

The node uses a small self-contained client rather than the chat panel's provider classes because exported scripts must run without EdgeWeave; the environment variables and default models are the same.

Embedding methods

Method	Needs	Matches on
tfidf (default)	nothing — scikit-learn, offline, deterministic	shared words
ollama	Ollama running + ollama pull nomic-embed-text (~270 MB)	meaning
openai	OPENAI_API_KEY (text-embedding-3-small)	meaning
auto	—	Ollama's nomic-embed-text if it is pulled, otherwise TF-IDF

Chat-only Ollama models (e.g. gemma3) can't produce embeddings; the node says so and suggests an embedding model.

Python export

Every node exports: the same ai_* functions that run in the app are copied into the script, so export and runtime can't drift. Relative folders in Load Documents resolve from the working directory when the script runs (in the app they also fall back to the EdgeWeave root and the active project).

While building this family we fixed an exporter bug: {tokens} in a node's expr were filled by chained str.replace, so a field value that itself contained {name} text (a Prompt Template's {context}) was rewritten by a later substitution. PythonGenerator._expand_expr now substitutes all tokens in a single pass.