In-app AI chat assistant "any" for finding and creating plots

[MarkusNeusinger]

Idea

Add an AI chat assistant called "any" that lives in a small floating/overlay chat
window on anyplot.ai. Users can describe their data or goal in natural language, and
"any" helps them in three modes:

1. Find — recommend matching specs from the gallery
2. Navigate — drive the SPA to specs, libraries, filters
3. Create — help propose / generate a new plot when nothing fits

User experience

• Persistent floating chat bubble (collapsible overlay, doesn't block content)
• User types something like "I have a time series with two categories and want to
  compare trends"
• "any" suggests relevant specs (e.g. line-timeseries-multi, area-stacked-temporal)
  with a preview thumbnail and a one-line rationale, and can navigate to them
• Follow-up questions stay in context: "show me the plotly version", "something
  more minimal", "compare to bar chart alternatives"
• If nothing fits: "I don't see a good match — want me to draft a new spec?"

Capabilities

Finding

• Semantic match against specification.md + tags
• Filter / refine via natural language (library, domain, complexity)

Navigating

• Programmatic SPA routing (React Router) to spec / library / catalog pages
• Apply / clear catalog filters
• Switch library tab inside SpecDetailView
• Copy code, open interactive preview

Creating

• Conversational spec drafting: ask clarifying questions about data shape, goal,
  audience
• Produce a draft specification.md + suggested tags
• Submit as a GitHub Issue with spec-request label (so the existing
  spec-create.yml pipeline takes over) — never bypass the workflow
• Optionally attach the user's data sample / sketch as context
• Show the user a link to the resulting issue / PR for follow-up

Architecture sketch

Framework: Google ADK (Agent Development Kit). Preferred over alternatives
because of its clean tool/agent model, streaming support, and good developer
ergonomics from prior experience.

• Agent runtime: ADK agent hosted alongside the FastAPI backend (separate
  process or sub-app), exposes a streaming chat endpoint consumed by the SPA
• Model: Claude (consistent with the rest of the platform — code generation
  and quality review already use Claude); ADK supports multi-model, so this
  can be swapped if needed
• Tool surface (frontend, called via tool-use → SPA action bridge):
  navigate, applyFilter, selectLibrary, copyCode, openInteractive
• Tool surface (backend, native ADK tools): searchSpecs, getSpec,
  draftSpec, createSpecRequestIssue
• Memory: per-session context via ADK session state (data shape, prior
  preferences); no cross-session persistence in v1
• Frontend: small React component (floating bubble + chat panel) that
  consumes the streaming endpoint and dispatches frontend tool calls into
  the existing SPA (router + filter hooks)

Open questions

• Deployment shape: same Cloud Run service as the API, or separate service?
• Auth / rate limiting (anonymous vs. signed-in)
• Free (Core) vs. premium gate per docs/concepts/vision.md
• Telemetry via Plausible: recommendation acceptance, create-flow conversion
• Should "create" mode also support uploading a data sample (ties into the
  premium "your data" feature in the vision doc)?
• Relationship to the existing MCP server (api/mcp/) — share tools or
  keep separate?

Related

• docs/concepts/vision.md — natural-language search, MCP server, "your data"
• api/mcp/ — existing MCP server (potential tool-sharing target)
• Existing spec-creation pipeline (spec-create.yml) — must remain the single
  path for new specs; "any" only drafts and submits the request issue