AI and RAG

Language-model and retrieval nodes. See also the AI / LLM / RAG guide.

8 nodes. Right-click any node in the editor to read this documentation in the app.

AI / LLM

💬 LLM Chat

id ai_llm_chat · AI / LLM · Python export: yes

Send a prompt to a large language model and output its reply. Providers: OpenAI, Anthropic Claude, Google Gemini, Ollama (local, no key) or any OpenAI-compatible server. Leave Model blank for the provider's default (Ollama: your first local model). API keys come from the project .env (OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY, OLLAMA_HOST, OPENAI_COMPATIBLE_BASE_URL/_API_KEY/_MODEL) — never from the graph, and exported scripts read them from the environment too. Each run makes a real (possibly billed) API call.

Inputs

Port	Type	Description
prompt	str	The user prompt (overrides the Prompt field) — e.g. from Prompt Template. Overrides the Prompt field when connected.
system	str	System instructions (overrides the System field). Overrides the System field when connected.

Outputs

Port	Type	Description
reply	str	The model's reply text.

Fields

Field	Type	Default	Choices
Provider	select	openai	openai, anthropic, gemini, ollama, openai_compatible
Model (blank = default)	text
System	textarea	You are a helpful assistant.
Prompt	textarea
Temperature	float	0.2
Max output tokens (0 = default)	int	0

🧩 Prompt Template

id ai_prompt_template · AI / LLM · Python export: yes

Build a prompt from a template. {question} and {context} are filled from the matching inputs, and any other {name} from the keys of the 'values' dict input. Placeholders with no value are left as written, so literal braces (e.g. in code) are safe.

Inputs

Port	Type	Description
question	str	Fills {question} (overrides the Question field). Overrides the Question field when connected.
context	str	Fills {context} — e.g. from Format Context.
values	dict	Optional dict: each key fills the {key} placeholder.

Outputs

Port	Type	Description
prompt	str	The filled-in prompt.

Fields

Field	Type	Default	Choices
Template	textarea	Answer the question using only the context belo…
Question	text

AI / RAG

🗂️ Build Vector Index

id ai_build_index · AI / RAG · Python export: yes

Index chunks for similarity search. tfidf (default) works offline with no key or download and matches on shared words; ollama / openai use embedding models that also match meaning (see Embed Text for the details); auto uses Ollama's nomic-embed-text when it is pulled, otherwise TF-IDF. The index is a plain dict, so it can be saved or exported like any other value.

Inputs

Port	Type	Description
chunks	list[dict] | list[str]	From Split Text (or a list of strings).

Outputs

Port	Type	Description
index	VectorIndex	Dict: method, model, chunks, matrix (+ the TF-IDF vectorizer).

Fields

Field	Type	Default	Choices
Method	select	tfidf	tfidf, auto, ollama, openai
Model (blank = default)	text

🧮 Embed Text

id ai_embed_text · AI / RAG · Python export: yes

Turn texts into vectors (one row per text, unit length) — e.g. to cluster or plot them with the ML nodes. tfidf: offline, fitted on these texts. ollama: a local embedding model (default nomic-embed-text; run ollama pull nomic-embed-text). openai: text-embedding-3-small (needs OPENAI_API_KEY). auto: Ollama's model when it is pulled, otherwise TF-IDF.

Inputs

Port	Type	Description
texts	list[str] | list[dict]	Strings, or chunks/documents (their 'text' is used).

Outputs

Port	Type	Description
embeddings	ndarray	Array of shape (n_texts, dimensions).

Fields

Field	Type	Default	Choices
Method	select	tfidf	tfidf, auto, ollama, openai
Model (blank = default)	text

📋 Format Context

id ai_format_context · AI / RAG · Python export: yes

Join retrieved chunks into one numbered context string for a prompt, optionally labelled with each chunk's source so the model can cite [1], [2]… Stops adding chunks once the character limit is reached (0 = no limit).

Inputs

Port	Type	Description
hits	list[dict] | list[str]	From Retrieve.

Outputs

Port	Type	Description
context	str	The formatted context text.

Fields

Field	Type	Default	Choices
Label with sources	checkbox	true
Max characters (0 = no limit)	int	6000

📚 Load Documents

id ai_load_documents · AI / RAG · Python export: yes

Read every text file in a folder that matches a glob pattern (comma-separate several, e.g. '**/.md, **/.txt'; ** searches subfolders). Relative folders are resolved from the EdgeWeave root, like other file nodes. Binary or non-UTF-8 files are skipped.

Inputs

Port	Type	Description
folder	str	Folder path (overrides the Folder field). Overrides the Folder field when connected.

Outputs

Port	Type	Description
documents	list[dict]	One {'source', 'text'} dict per file (source = path relative to the folder).

Fields

Field	Type	Default	Choices
Folder	text
Pattern	text	**/*.md
Max files	int	200

🔎 Retrieve

id ai_retrieve · AI / RAG · Python export: yes

Find the chunks most similar to a query (cosine similarity) — the 'R' in RAG. Scores run from 0 (unrelated) to 1 (identical).

Inputs

Port	Type	Description
index	VectorIndex	From Build Vector Index.
query	str	The question to search for (overrides the Query field). Overrides the Query field when connected.

Outputs

Port	Type	Description
hits	list[dict]	Best-first {'rank', 'score', 'source', 'chunk', 'text'} dicts.

Fields

Field	Type	Default	Choices
Query	text
Results	int	4

✂️ Split Text

id ai_split_text · AI / RAG · Python export: yes

Split documents into chunks small enough to retrieve and fit in a prompt. 'paragraphs' packs blank-line-separated paragraphs up to the chunk size; 'markdown_sections' starts a new chunk at every heading; 'characters' uses fixed windows. Overlap repeats the end of one chunk at the start of the next so sentences aren't lost at the seams.

Inputs

Port	Type	Description
documents	list[dict] | list[str] | str	From Load Documents, or plain strings.

Outputs

Port	Type	Description
chunks	list[dict]	One {'source', 'chunk', 'text'} dict per chunk.

Fields

Field	Type	Default	Choices
Split by	select	paragraphs	paragraphs, markdown_sections, characters
Chunk size (characters)	int	800
Overlap (characters)	int	100