repo-pulse

AI Chat

RepoPulse includes a conversational AI panel that lets users ask natural-language questions about any analysis result — a single repo, a multi-repo comparison, an org inventory, or a full org health run.

How it works

The chat panel appears as a slide-up tray at the bottom of every analysis view. It is scoped to the currently displayed data: the model only sees the serialized analysis context (metrics, scores, recommendations) for whatever is on screen, never raw GitHub tokens or unrelated data.

Each message exchange follows this flow:

1. The frontend serializes the visible analysis into a JSON context block.
2. It sends that context + the conversation history to POST /api/chat.
3. The API route authenticates the user, enforces quotas, resolves a provider + model, and streams a response via the Vercel AI SDK.
4. The streamed response is displayed incrementally in the chat bubble.

The user can stop a response mid-stream using the square stop button that appears while the model is typing.

Providers

The chat supports five providers. The user selects a provider and model in the chat settings panel; the choice is persisted in localStorage.

Provider	Models available
Anthropic	claude-haiku-3-5 (fast), claude-sonnet-4-5 (quality)
OpenAI	gpt-4o-mini (fast), gpt-4o (quality)
Google	gemini-2.0-flash (fast), gemini-2.5-pro (quality)
Groq	llama-3.1-8b-instant (fast), llama-3.3-70b-versatile (quality)
OpenRouter	meta-llama/llama-3.1-8b-instruct (fast), anthropic/claude-3.5-sonnet (quality)

Free tier

If the server has at least one AI provider key configured (ANTHROPIC_API_KEY, OPENAI_API_KEY, GOOGLE_API_KEY, or GROQ_API_KEY), signed-in users get 5 free chats per GitHub login per calendar day (UTC midnight reset). The first key found in that priority order is used for free-tier requests.

• The remaining count is shown as colored dots in the chat panel.
• Once exhausted, the panel prompts the user to enter their own API key.
• Failed requests (provider errors, network timeouts) do not consume quota — the counter is decremented on error.

If no server key is configured, the free tier is unavailable and users must always supply their own key.

Bring your own key

Any user can enter their own API key in the chat settings panel. With an own key:

• Provider and model selection are fully unlocked.
• Daily quota does not apply.
• The key is sent to POST /api/chat in the request body and forwarded to the selected provider. It is never logged, stored, or persisted server-side.

Context scoping

The context sent to the model is trimmed to the most relevant fields:

• Repos mode: per-repo metrics snapshot (stars, commits, PRs, issues, contributors, security/doc/license results).
• Org mode (inventory phase): repo list with stars, language, topics, push date, license.
• Org mode (analysis done): org-level panel rollups (security, bus-factor, documentation coverage, etc.) plus a per-repo status table.

Context is serialized as a JSON block inside the system prompt. The model is instructed to answer strictly from the provided data and not to invent metrics.

Conversation history

Up to 10 turns (20 messages) of history are sent with each request. Older messages are trimmed from the front. A visual divider is inserted in the UI whenever the underlying analysis context changes (e.g., user filters to different repos) so it’s clear which data a prior answer was based on.

Starter chips

The panel surfaces suggested questions as clickable chips:

• Repos: “What are the biggest health risks?”, “Which repo is most active?”, “How does security compare across these repos?”, etc.
• Org (inventory): “Which repos have the most stars?”, “How many repos use each language?”, etc.
• Org (analysis): “Which repos need the most attention?”, “What are the top recommendations?”, etc.

Chips disappear once the first message is sent.

Server setup

See docs/DEPLOYMENT.md for the environment variables required to enable the free tier.

This site is open source. Improve this page.