AI Providers

Spark AI, SparkFeed’s built-in chat that lets you ask questions about your feeds, is powered by the Vercel AI SDK. You choose which AI provider to use by setting the appropriate API key. No code changes required.

Provider priority

SparkFeed checks for API keys in this order and uses the first one found:

OPENAI_API_KEY  →  AI_GATEWAY_API_KEY  →  OPENROUTER_API_KEY

Set exactly one key (or multiple, in which case the first match wins). If no key is set, the AI chat panel is disabled at runtime.

Option 1: OpenAI

Best for most users. Straightforward, well-documented, and the default small/large model pairing is cost-efficient.

# .env.local
OPENAI_API_KEY=sk-...

Get a key at platform.openai.com/api-keys.

Default models used:

• Small (fast, cheap): gpt-4o-mini
• Large (high-quality): gpt-4o

Option 2: Vercel AI SDK Gateway

Recommended for Vercel users

The Vercel AI Gateway is a unified proxy that gives you access to multiple model providers through a single API key, with built-in usage tracking.

# .env.local
AI_GATEWAY_API_KEY=...

Get a key in your Vercel dashboard → AI → Gateway.

Default models used:

• Small: openai/gpt-4o-mini
• Large: openai/gpt-4o

Option 3: OpenRouter

Access 200+ models from a single API key, including Claude, Gemini, DeepSeek, Llama, and more.

# .env.local
OPENROUTER_API_KEY=sk-or-...

Get a key at openrouter.ai/keys.

Default models used:

• Small: openai/gpt-4o-mini
• Large: openai/gpt-4o

Model selection

The model registry lives in src/config/ai-models.json. It defines available models per provider with labels, tiers (small / large), and context windows. The default model for each provider is picked automatically based on the provider detected at startup.

Override the model

Set AI_MODEL to any model ID supported by your chosen provider:

# .env.local
OPENAI_API_KEY=sk-...
AI_MODEL=gpt-4o                     # override: use large model by default

AI_GATEWAY_API_KEY=...
AI_MODEL=anthropic/claude-3-5-haiku  # Claude via gateway

OPENROUTER_API_KEY=sk-or-...
AI_MODEL=deepseek/deepseek-chat-v3   # DeepSeek via OpenRouter

Avoid very expensive models

Models like Claude Opus or GPT-4.5 are significantly more expensive and offer diminishing returns for summarizing RSS articles. Stick to gpt-4o / gpt-4o-mini or equivalent mid-tier models unless you have a specific reason.

Available models reference

The full model registry is at src/config/ai-models.json. Supported providers and their included models:

OpenAI

Model ID	Tier	Context
gpt-4o-mini	Small	128k
gpt-4o	Large	128k

Vercel AI Gateway / OpenRouter: Claude

Model ID	Tier	Context
anthropic/claude-3-5-haiku-20241022	Small	200k
anthropic/claude-3-7-sonnet-20250219	Large	200k

Vercel AI Gateway / OpenRouter: DeepSeek

Model ID	Tier	Context
deepseek/deepseek-chat-v3	Small	64k
deepseek/deepseek-r1	Large	64k

OpenRouter: MiniMax & Kimi

Model ID	Tier	Context
minimax/minimax-01	Small	1M
moonshotai/kimi-k2	Large	128k

How the AI chat works

When you send a message in the Spark AI panel:

1. The client sends the message + your workspaceId to /api/ai
2. The server queries your articles for context (latest articles from your feeds)
3. The message + article context is streamed to the selected model
4. The response streams back to the UI with source citations

The AI only has access to articles stored in your database. It does not fetch the internet on your behalf.

No key set?

If no AI key is configured, the Spark AI panel remains visible but chat is disabled at the API level. Set any key from the options above to activate it.