node-red-contrib-mistral

Node-RED nodes for the Mistral AI API — chat completions (with vision and streaming), embeddings, and OCR.

Table of contents

• Install
• Quick start
• Nodes
  • mistral-config
  • mistral-chat
  • mistral-embeddings
  • mistral-ocr
• Example flows
• Recipes
  • RAG with embeddings
  • Streaming to a dashboard
• License

Install

From the Node-RED editor:

1. Open the palette manager (hamburger menu → Manage palette).
2. Switch to the Install tab.
3. Search for node-red-contrib-mistral and click Install.

Or from the command line in your Node-RED user directory (typically ~/.node-red):

npm install node-red-contrib-mistral

Then restart Node-RED. Requires Node.js ≥ 18 and Node-RED ≥ 3.0.

Quick start

1. Drag a mistral chat node onto your flow.
2. Double-click it. In the Server field, click the pencil to create a new mistral-config entry.
3. Paste your Mistral API key (get one at console.mistral.ai) and save.
4. Wire an inject node (string payload) → mistral chat → debug.
5. Deploy, click inject, watch the response in the debug sidebar.

The config node is shared across all Mistral nodes — set it up once, reuse it everywhere.

Nodes

mistral-config

Holds the API key (stored encrypted in Node-RED's credentials store) and base URL. Not visible in the palette — it's only accessible from the editor panel of action nodes via the Server dropdown.

| Field | Default | Notes | |---|---|---| | API Key | — | Required. Stored encrypted. | | Base URL | https://api.mistral.ai/v1 | Override for proxies or self-hosted gateways. |

mistral-chat

Calls POST /chat/completions. Supports text, vision (image inputs), and streaming.

Inputs

| Property | Type | Description | |---|---|---| | msg.payload | string | ChatMessage[] | User prompt as string, or full conversation array. | | msg.image | string | string[] | Optional. Image URL(s) or data: base64. Requires a vision model (e.g. pixtral-large-latest). | | msg.model | string | Optional. Override the configured model. | | msg.temperature | number | Optional. | | msg.max_tokens | number | Optional. | | msg.stream | boolean | Optional. Force streaming on/off for this call. |

Outputs (non-streaming)

| Property | Type | |---|---| | msg.payload | string — assistant reply | | msg.mistral | full API response (id, model, choices, usage) | | msg.complete | true |

Outputs (streaming) — one msg per chunk plus a final summary:

| Property | Chunks | Final | |---|---|---| | msg.payload | delta text | full accumulated text | | msg.complete | false | true | | msg.mistral | — | last SSE event with usage |

mistral-embeddings

Calls POST /embeddings to vectorize text. Default model: mistral-embed (1024 dims).

Inputs

| Property | Type | |---|---| | msg.payload | string | string[] | | msg.model | optional string override |

Outputs

| Property | Type | |---|---| | msg.payload | number[] if single input, number[][] if array input — aligned to input order | | msg.mistral | full API response |

mistral-ocr

Calls POST /ocr to extract markdown from PDFs and images. Default model: mistral-ocr-latest.

Inputs

| Property | Value | |---|---| | msg.payload | URL string (https or data:), OR an object { document_url } / { image_url } / { url } |

The node auto-detects document vs image from the URL extension by default. You can force a type in the editor.

Outputs

| Property | Type | |---|---| | msg.payload | concatenated markdown across all pages | | msg.pages | per-page array { index, markdown, images?, dimensions? } | | msg.mistral | full API response |

Example flows

Importable JSON flows are in examples/:

• test-flow.json — exercises chat, streaming, embeddings, and OCR.
• vision-flow.json — minimal Pixtral image-description flow.
• rag-flow.json — full RAG pattern (index → embed query → retrieve → answer).

Import in Node-RED: hamburger menu → Import → select a file to import (or paste the JSON content).

⚠️ The mistral-config ID in each example may collide with an existing config in your flow. Node-RED will prompt to Replace or Copy. Choose Replace to reuse your existing config (with your saved API key), or Copy and re-attach manually.

Recipes

RAG with embeddings

The rag-flow.json example shows the canonical pattern in 6 steps:

1. Index — inject array of documents → mistral-embeddings → function node stores {text, vec} pairs in flow.context.
2. Query — inject question → mistral-embeddings (single string → single vector).
3. Retrieve — function node computes cosine similarity vs the stored vectors and keeps the top-K.
4. Build prompt — same function node concatenates retrieved chunks into a context block.
5. Generate — mistral-chat answers from the context with a system prompt enforcing context-grounding.
6. Debug — final answer + the retrieved chunks for traceability.

For production, replace the in-memory flow.context store with a proper vector database (Qdrant, pgvector, Pinecone, Chroma) — the only nodes you need from this package are mistral-embeddings and mistral-chat.

Streaming to a dashboard

For a live-typing UI:

inject → mistral-chat (stream=true) → function (route by msg.complete) → ui-template (append chunk)
                                                                       → ui-template (mark complete)

Filter chunks (msg.complete === false) to your "append" UI handler, and the final message (msg.complete === true) to a "done" handler that displays usage stats.

Contributing

Issues and pull requests welcome on the project repository.

License

MIT © Damien Lachambre