[M5] LLMFrame: surface current context into Claude prompt with suggested actions

[Korede-TA]

Goal

A frame that assembles the user's current selection and corpus context into a structured prompt and offers a set of short, actionable suggested prompts — enabling the user to move from exploration to synthesis in one click.

Context assembly

The frame builds a context block from live store state:

• Selected nodes: deriveLabel + full node.text (truncated at ~400 tokens total across all selected)
• Temporal range: if a date filter or selection has a time span, include "Documents spanning [year] – [year]"
• Geographic focus: top location names from selected nodes' geo annotations
• Key entities: top person/org names from selected nodes' entity annotations
• Current level + scope: e.g. "Viewing 23 chunks within document 'Lake Chad Report'"

Suggested prompts

Show 4–6 short prompt chips that are dynamically generated based on what's in context. Examples based on selection type:

• Single document selected: "Summarize this document", "What are the key claims?", "List the locations mentioned"
• Multiple chunks selected: "What themes connect these passages?", "Identify contradictions across these excerpts"
• Geographic selection: "What events happened in [top location]?", "Describe the geographic pattern"
• Temporal selection: "Summarize what changed between [year] and [year]"
• Entity-heavy selection: "What is the role of [top entity] across these documents?"

Execution

• Clicking a suggested prompt (or editing + submitting a custom one) sends the context + prompt to the Claude API via akb serve backend endpoint (POST /llm)
• The akb serve endpoint calls anthropic.messages.create with the assembled context as a system block and the prompt as the user turn
• Response streams back and is rendered in the frame as markdown
• Response includes a "Copy" button and a "Save to notes" action (appends to a session notes store)

UI layout

• Top section: read-only assembled context preview (collapsible, shows token count)
• Middle: 4–6 suggested prompt chips (regenerate button)
• Bottom: editable prompt textarea + Send button
• Response area: streaming markdown output below

Registration

• Register 'llm' in frame registry and add to ADDABLE_FRAMES
• Backend: add POST /llm to akb serve (FastAPI, streams SSE)

Acceptance criteria

• Suggested prompts update within 200ms of selection change
• Prompt chips reflect the actual content of the selection (not generic defaults when context is available)
• Response streams visibly (not a single delayed render)
• Works offline in degraded mode: shows context + prompts but disables Send with "requires akb serve" message