ai-analyst

name

ai-analyst

description

Python sidecar (LangGraph indicator/pattern/trend/decision agents + TA-Lib + vision LLM) reached over HTTP/JSON via IAiAnalystClient; hot-swappable Null↔Http via IOptionsMonitor. Use when touching tools/python-ml/, the IAiAnalystClient seam in Core/Infrastructure, the AiAnalystEnricher in the notification pipeline, the AI Market Analyst window, or debugging "AI Analyst unavailable" / HTTP 504 / 500 errors.

AI Analyst

A multi-agent LLM analyst lives in a Python sidecar so the C# build stays hermetic — no Python, no TA-Lib C extension in the .NET tree. The WPF app talks to it over http://127.0.0.1:<port>/analyst/run. Bring your own API key (OpenAI / Anthropic / Qwen / MiniMax).

Architecture

C# side (src/TradingTerminal.Core/AiAnalyst/ + src/TradingTerminal.Ai/Analyst/):

• IAiAnalystClient — bool IsAvailable + Task<AnalystReport> RunAsync(AnalystRequest, ct).
• NullAiAnalystClient — always returns "unavailable". Default when sidecar isn't enabled.
• HttpAiAnalystClient — POSTs to the sidecar. Uses named HttpClient (HttpAiAnalystClient.HttpClientName) with a 2 min outer ceiling; per-call timeout via AiAnalystOptions.TimeoutSeconds (default 60s).
• DispatchingAiAnalystClient — the registered IAiAnalystClient. Reads IOptionsMonitor<NotificationsOptions>.CurrentValue.AiAnalyst.Enabled on every call → hot-swaps Null↔Http without restart.
• AiAnalystEnricher : INotificationEnricher — appends AI Analyst: Long (conf 72%, R:R 2.10) — Bull Flag to every Telegram/Discord signal when enabled. Failures swallowed silently → original notification still ships.

Python side (tools/python-ml/):

• daxalgo_ml/ package, pyproject.toml, daxalgo_ml.spec (PyInstaller).
• Build: pip install -e .[dev] then pyinstaller daxalgo_ml.spec → dist/daxalgo-ml.exe. Drop into tools/python-ml/bin/.
• Run: $env:DAXALGO_ML_PORT = "8765"; .\bin\daxalgo-ml.exe.
• Binds 127.0.0.1 only. Never 0.0.0.0.
• Four-agent LangGraph:
  1. Indicator — TA-Lib RSI / MACD / ATR / EMA + text LLM regime summary. NumPy-only fallback if TA-Lib is missing (so tests don't require it).
  2. Pattern — matplotlib candlestick PNG → vision LLM scored against a 16-pattern classical catalog.
  3. Trend — fit upper/lower linear channel on highs/lows with one round of 2-sigma outlier rejection → annotated PNG → vision LLM for trend regime.
  4. Decision — strict-JSON verdict (long/short/no_call), up to 3 retries on invalid JSON.

Response: full AnalystReport with base64-encoded PNG charts the WPF view binds directly to Image controls.

Settings + secret handling

Configured under Tools → Settings → Notifications → AI Market Analyst:

• Enabled toggle.
• Endpoint (http://127.0.0.1:8765 default).
• Provider (OpenAI / Anthropic / Qwen / MiniMax).
• API key — DPAPI-encrypted in %LOCALAPPDATA%\DaxAlgo Terminal\notifications.json. Never plain-text. Never in appsettings.json. Scope: current user, current machine.
• Text model id + vision model id.
• Per-notification enrichment toggle (separate from Ollama; both can run).

When something breaks

• "AI Analyst unavailable" in the pane. Hit http://127.0.0.1:<port>/healthz in a browser. Not returning {"status":"ok"} → sidecar not running OR wrong port in Settings.
• HTTP 504 — analyst run timed out. First vision call is often slow. Retry once. If chronic, raise AiAnalystOptions.TimeoutSeconds (don't exceed the outer 2 min ceiling).
• HTTP 500 — API key for provider 'X' is empty. User skipped Save after pasting. Re-paste + Save.
• Wrong-paste API key. DPAPI round-trip will fail → re-paste + Save.
• Sidecar console: missing TA-Lib. Use the Gohlke prebuilt wheel for Windows (pip install TA-Lib fails for some Python versions); or rely on the NumPy fallback for unit tests.
• AiAnalystEnricher silently dropped a line. That's by design — sidecar failure must not break the notification path. Check Serilog for the failure trace.

Hard rules

• Sidecar binds 127.0.0.1 only. Loopback + process-bound lifetime = same trust boundary as the WPF process. Binding to 0.0.0.0 would expose API keys + arbitrary chart rendering to anything on the LAN.
• No Python or TA-Lib in the C# build. The seam stays HTTP/JSON subprocess. If you find yourself adding <PackageReference Include="Python.Net" /> — stop.
• No "Place this trade" button on the analyst pane. Signal-mode only by rule. The analyst proposes; the human disposes.
• Enricher must never throw or block. Sidecar down = original notification unchanged, no exception surfaces.
• Hot-swap via IOptionsMonitor.CurrentValue — read fresh on every call. Don't cache the Enabled flag.

Reference reads

• src/TradingTerminal.Core/AiAnalyst/IAiAnalystClient.cs — seam.
• src/TradingTerminal.Ai/Analyst/AiAnalystServiceCollectionExtensions.cs — DI.
• src/TradingTerminal.Ai/Analyst/HttpAiAnalystClient.cs — HTTP path.
• src/TradingTerminal.Ai/Analyst/AiAnalystEnricher.cs — enricher.
• tools/python-ml/daxalgo_ml/ — sidecar source.
• docs/ai-analyst.md + docs/polyglot.md — user-facing + architecture prose.

See also: add-notifier, [[project-ai-analyst]] (memory).