AI and Chat

AI and Chat

Everything AI-related talks to a local Ollama server. Nothing is sent to cloud LLM APIs.

Packages

Piece	Location
HTTP client	internal/ai/ollama.go
Bindings	ai_api.go, ai_commit.go, chat_api.go, codectx_api.go
Retrieval	internal/codectx
Persistence	internal/chatstore
UI settings	frontend/src/ai.js (localStorage)

Ollama client

• Default http://localhost:11434, override via argument or OLLAMA_HOST
• Long HTTP timeout (minutes) for local generation
• Used for: task enrich, freeform chat, commit messages, project descriptions

Code context (codectx)

• Walks the project (skips heavy dirs), chunks source, builds a BM25 index — no embedding model required
• Overview (tree, manifests, scripts) is cheap and often included in prompts
• Search returns top chunks pasted into the system prompt
• Progress events: codectx:progress, codectx:done
• Index is per-project and removed on DeleteProject

Chat store

• Path: ~/.jumpstart/chats/<safeProjectID>.json
• Kept out of config.json so chat traffic does not rewrite the project file
• Cap: 50 sessions per project (oldest dropped)
• Messages may store structured stories JSON and sources file paths for UI previews

UI surfaces

1. TaskDetailModal — “Fill with AI” → OllamaEnrichTask
2. ChatDock — story assistant; expand to fullscreen; SendChatMessage
3. CommitBox — OllamaGenerateCommitMessage using gitops commit context
4. ProjectModal — optional AI project description

Privacy constraints

Analytics may record model_family, param_size, lengths, latency, success — never prompts, replies, paths, or raw model tags. Use TrackModelSelected so Go parses the tag safely.

Changing prompts / retrieval

• Prompt assembly: internal/codectx/prompt.go and enrich paths in ai_api.go
• Chunking / walk limits: chunk.go, walk.go
• Keep retrieval local and deterministic unless you intentionally add a new dependency (and document the download UX)