AI Agent Chat Assistant for Portfolio Analysis #6434
Replies: 3 comments 4 replies
|
Follow-up: the scoped implementation is now open for review as PR #7444: It stays within the proposed MVP: an inline Analysis-page chat using the existing Vercel AI SDK/OpenRouter configuration, three fixed read-only portfolio tools, active filters/date range, session consent, no persistence, and no write operations. It also includes user/global eligibility gates, a fail-closed 5 requests/minute user limit, bounded messages/output/steps/time, redacted tool payloads, disconnect cancellation, and restricted model-output rendering. Validation completed:
Feedback on the scope and safety boundaries would be very welcome. |
|
@richardchen874-sys Addressed in #7444: consent is now bound to a normalized signature of the displayed portfolio filters, date-range selection, and model. Changing that scope stops any active response, clears the conversation, invalidates the prior consent, and returns the user to the disclosure for re-consent; filter reordering and label-only changes remain equivalent. Each request still submits its exact filters and date-range selector, so the server-side tools for that stream remain bound to the same request-local scope. This is intentionally scope-bound consent rather than a transactionally frozen portfolio-data snapshot. |
|
Thanks for your proposal, @shoyu-ramen! We are still evaluating our overall approach to AI, with a focus on privacy. We are undecided on whether a built-in chat interface makes sense for Ghostfolio or if it is better to let users connect external AI tools of their choice. |
Uh oh!
There was an error while loading. Please reload this page.
Summary
I'd like to propose adding an opt-in AI chat assistant to Ghostfolio that helps users analyze their portfolio through natural language conversation.
What it does
ANTHROPIC_API_KEYis set. Zero impact on existing functionality when disabled.Technical approach
@anthropic-ai/claude-agent-sdk) for the agent runtimegf-agent-chat) following existing UI patternsWhat's NOT included
New dependencies
@anthropic-ai/claude-agent-sdk@opentelemetry/*(optional, for observability)zod(schema validation for tool parameters)Security
ANTHROPIC_API_KEYI have a working implementation on my fork and would like to submit a PR. Happy to discuss the approach and adjust based on feedback.
cc @dtslvr
All reactions