Update Logs

This page tracks the history of updates and new features for Agentic FinSearch.

Version 0.16.0

XBRL Validation Pipeline: A three-stage tagging → retrieval → matching pipeline that checks numerical claims in agent responses against SEC XBRL filings and surfaces per-claim verdicts. See XBRL Validation Pipeline.
Validate Button: User-triggered, deterministic validation. Appears in the response toolbar only when the response emitted at least one supported claim; clicking it runs the pipeline and renders inline mismatch marks plus a Ground Truth section in the Sources popup.
Three Layer 1 axioms: Balance-sheet equality (A = L + E with TempEquity handling for filers with non-controlling interests), gross margin, and current ratio — evaluated against three pre-loaded SEC filings (AAPL FY2023, MSFT FY2023, TSLA FY2023). Wider coverage will arrive with the upcoming cloud-hosted SEC XBRL Filing Tree.
``Not Applicable`` verdict: Deterministic detection of metrics that are undefined for a filer (e.g., current ratio on an unclassified balance sheet from a bank or REIT), distinct from Skipped, so the audit log stays honest.
Ground Truth provenance: The XBRL filing path used by each validation pass is bridged into the existing Sources popup via the claim registry, so users see exactly which filing was consulted per claim.
Installation pages regrouped: All installation guides now live under a single Installation section in the navigation to keep the top-level toctree compact.

Query Planner & Skills: Heuristic-based skill routing that constrains MCP tools per query type for faster, more focused responses. Six skills: summarize_page, stock_fundamentals, options_analysis, financial_statements, technical_analysis, web_research.
Hallucination Mitigation: Improved accuracy for financial data retrieval benchmarks.

OpenAI-Compatible API: The /v1/chat/completions API is now fully supported for external consumers.

OpenAI-Compatible REST API: New /v1/chat/completions and /v1/models endpoints with Bearer token authentication, rate limiting, and three modes (thinking, research, normal).
API Deployment Readiness: Full production deployment support with CORS, rate limiting, health checks, and environment-based configuration.
Research Engine Refinements: Improved streaming integration with the research engine for both API and frontend call paths.

Deep Research Mode: Multi-step research engine with QueryAnalyzer, ResearchExecutor, GapDetector, and Synthesizer for complex financial queries.
Iterative Research Loop: run_iterative_research orchestration that decomposes questions, executes sub-queries in parallel, detects gaps, and synthesizes results.
Research Config: Configurable research parameters in models_config.py (planner model, research model, max iterations, parallel search).

TradingView MCP Server: Custom MCP server for fetching technical analysis data and market screener results.
Gemini Integration: Google Gemini models (gemini-3-flash-preview) added as a foundation model provider with 1M token context support.
Prompt Architecture: Site-specific prompt templates in prompts/sites/ for Yahoo Finance, SEC EDGAR, and TradingView.
Enhanced MCP Prompt: Improved Yahoo Finance MCP prompt engineering for more accurate data retrieval.

Yahoo Finance MCP: Wrote a custom Yahoo Finance MCP server. Agent may now fetch data from Yahoo Finance much more accurately and precisely.
Performance Improvements: Enhanced the custom tree-based scraper to be more stable.