AI Features
Velocity includes built-in AI capabilities powered by a provider-agnostic LLM gateway. Features range from smart search to autonomous agents, with usage tracking and a monthly credit allowance per seat.
Overview
AI features in Velocity are plan-gated and can be enabled or disabled per workspace. Every AI call is tracked for token usage and cost. A workspace's monthly AI allowance is its seats times the plan's credits per seat.
Highlights
AI Search
Hybrid keyword + semantic search across issues and documents. Understands meaning, not just keywords.
Smart Issue Creation
Describe an issue in natural language. AI extracts title, priority, team, labels, and project for you.
Available Features
| Feature | Description | Min Plan |
|---|---|---|
| AI Search | Semantic search across issues and documents using vector embeddings. Results are merged with keyword matches in a hybrid approach. See AI Search. | Pro |
| AI Triage / Smart Create | Parse natural language into structured issue fields (title, description, priority, labels, team, project). Constrained to valid workspace values. See Smart Issue Creation. | Business |
| AI Description Writer | Generate detailed issue descriptions from a brief title or summary. | Pro |
| AI Comment Drafting | Draft responses and updates for issue comments. | Pro |
| AI Summarization | Summarize long issue threads, projects, or documents into concise overviews. | Pro |
| AI Document Drafting | Generate full document drafts from a topic or outline. | Business |
| AI Release Notes | Auto-generate release notes from completed issues in a cycle or project. | Business |
| AI Agent | Autonomous agent that can research, plan, and execute multi-step tasks on your behalf. | Business |
| Bring Your Own Key | Use your own API keys for AI providers instead of Velocity’s built-in credits. | Enterprise |
AI PRD draftingai_prd_generate | Clarify an idea and review an editable PRD draft before saving. See the workflow guide. | Business |
AI requirement-to-story breakdownai_story_breakdown | Review proposed stories, acceptance criteria, INVEST checks and requirement coverage before creating stories. See the workflow guide. | Business |
AI story check & splitai_story_refine | Check INVEST quality or propose smaller stories; apply changes only after review. See the workflow guide. | Pro |
AI story-to-issue proposalsai_decompose | Review engineering tasks and acceptance-criterion coverage, then confirm issue creation. See the workflow guide. | Business |
Planning costs and controls
Manual PRDs and stories are available on every plan. AI planning actions require the minimum plan above, an enabled workspace feature and available credits. Requirement-to-story breakdown and story-to-issue proposals show planned calls and cost before generation. PRD drafting surfaces an estimate for larger requests after its clarification step; clarification can itself use AI credits. Estimates are previews, not a price guarantee.
Each follow-up draft, quality check or retry can consume credits. Inspect scope before running; story breakdown can stop when the budget runs out and reports partial results. Review proposed changes before saving, applying a split or creating work. AI drafting never approves a PRD.
Owners and admins can inspect per-feature and per-member usage in Settings → AI usage and the credit allowance in Billing. An authorized setAiCreditOverride API mutation can set a workspace allowance; platform administrators manage feature flags in the admin dashboard. These controls still apply to planning actions. See From idea to issues.
Plan Tiers & AI Credits
Each plan includes a monthly allowance of AI credits per seat (seats times the plan's credits). One credit is $0.01 of AI usage. Usage is tracked in real time and visible in Settings → Billing.
| Plan | AI Enabled | AI credits per seat | Concurrent Streams |
|---|---|---|---|
| Free | No | 0 | 1 |
| Pro | Yes | 200 | 3 |
| Business | Yes | 500 | 5 |
| Enterprise | Yes | 1,500 | 10 |
LLM Gateway
Velocity uses a provider-agnostic LLM gateway that supports multiple AI providers. The gateway automatically routes requests through available providers and handles failover if one is unavailable. Platform AI uses the reviewed Vercel AI Gateway task policy; the direct-provider defaults below apply when those providers are configured.
Supported Providers
| Provider | Default Model | Used For |
|---|---|---|
| MiniMax | MiniMax-M3 | Direct completions and streaming |
| Kimi (Moonshot) | kimi-k2.6 | Direct completions and streaming |
| Anthropic | claude-sonnet-4-5-20250929 | AI Agent, complex tasks |
| OpenAI | gpt-4o-mini | Embeddings (text-embedding-3-small), fallback |
Configure keys for the providers' global APIs. MiniMax-M3 and Kimi-K2.6 default to non-thinking requests for routine work. Kimi's sampling settings are fixed by the model, so temperature overrides are omitted. Unsupported reasoning effort combinations are rejected before a request is sent. An explicit Kimi-K3 override uses low effort by default; high and max are also supported.
Usage is priced at each model's published rate, including cache reads and MiniMax-M3's higher rate above 512k input tokens. See the current MiniMax pricing and Kimi pricing.
Gateway Features
- Streaming responses — real-time token-by-token output with a typing cursor
- Concurrent stream limits — controlled per plan tier to prevent abuse
- Automatic retry — retries with exponential backoff on transient failures (429, 5xx)
- Provider fallback — if the primary provider fails, automatically tries the next
- Workspace overrides — workspaces can configure a preferred provider and model
- BYOK support — Enterprise workspaces can use their own API keys
Embedding Pipeline
Velocity generates vector embeddings for issues and documents to power semantic search. Embeddings use OpenAI’s text-embedding-3-small model (1536 dimensions) and are stored in PostgreSQL via the pgvector extension with HNSW indexing.
Embedding generation is fire-and-forget — when an issue or document is created or updated, an async request is dispatched to the embedding service. This means entity creation is never blocked by the embedding process. The embedding service deduplicates requests using content hashing to avoid unnecessary API calls.
Usage Tracking
Every AI API call is logged with:
- Feature name (search, triage, agent, embedding, etc.)
- Model and provider used
- Input and output token counts
- Cost in cents (calculated from per-model pricing)
- Duration in milliseconds
- User who triggered the request
Monthly aggregates are available via the aiUsage GraphQL query with per-feature breakdowns. The credit allowance is available via aiBudgetStatus.
Credit Allowance
When a workspace approaches its monthly AI credits:
- 80% used — a soft warning is shown to workspace admins
- 100% used — AI features are blocked until the allowance resets on the 1st (UTC), or the plan includes more credits
Workspace owners and admins can set a custom allowance with the setAiCreditOverride mutation.
Streaming & Concurrent Limits
AI features that use streaming (like the AI agent and description writer) deliver responses in real time with a typing cursor animation. Each user has a concurrent stream limit based on their workspace’s plan tier. If you exceed the limit, the request is rejected with a 429 status — wait for an existing stream to finish before starting a new one.
Feature Flag Overrides
Workspace administrators can override plan-tier AI access using feature flags. Feature flags use the ai:<feature_name> key convention (e.g., ai:ai_search) and can be managed in the admin dashboard under Feature Flags.
Feature flag overrides take precedence over plan tier defaults, allowing you to grant specific AI features to individual workspaces regardless of their plan.