FinSight AI
The AI research and reporting console for stock investment and research scenarios supports evidence retrieval, workflow tracking, financial indicators and RAG evaluation.
Support the online AI model in OPENAI format. AI provider fills in "openai-compatible", and other fields are filled in according to actual conditions. If online AI is not available, there are local rules. !!! The first boot will fail, and the application will resume after restarting. It will take a while to initialize the environment, please wait patiently. FinSight is an open source AI Agent backend platform for stock investment research scenarios. Its core capabilities include: evidence-driven AI research reports, recoverable Agent workflow, Redis Lua Single-flight concurrency control, PostgreSQL/pgvector hybrid retrieval, report versioned caching and RAG evaluation. This project is not a simple "tuning model interface" Demo, but focuses on demonstrating the back-end engineering capabilities behind the AI Agent: long-link task governance, idempotent scheduling, failure recovery, trusted caching, evidence tracking, and output quality evaluation. FinSight comes with an operational institutional investment and research console. This front-end is not a simple decorative page, but displays the investment research workflow, report cache, evidence chain, RAG evaluation and financial risk signals generated by the back-end. Why do you do this project? Many RAG projects stop at "retrieve a few paragraphs of text and ask LLM." FinSight is more concerned about the problems that need to be solved when an AI investment and research system is truly implemented: Long-running Agent workflows require clear state machines; Expensive tasks cannot be performed repeatedly in a multi-instance environment; AI calls require Single-flight to avoid request amplification and cache breakdown; AI reports cannot be cached just by prompt, but must be bound to data snapshots; RAG answers must be able to trace the source of evidence; The quality of AI output needs to be regressive and evaluated rather than relying solely on subjective feelings. core highlight module achieve content Agent workflow Break data collection, indicator recalculation, document indexing, company portraits, and AI research report generation into recoverable stages concurrency control Impotent key, repository layer createIfAbsent, Redis Lua single-flight lease, fencing token, local downgrade lock failure recovery Task state machine, phase trace, retry, Dead Letter, timeout takeover scheduler Trusted AI Cache contextHash, dataSnapshotHash, reportVersion, support Redis/PostgreSQL cache reuse retrieval link PostgreSQL JSONB, full-text search, pgvector, mixed recall, evidence deduplication evaluation system RAG hit rate, evidence coverage, answer coverage, risk of hallucination, conclusion consistency, confidence calibration, delay Demo Show Spring Boot API, static Dashboard, sample data flow, Actuator, Prometheus metrics
