Skip to content

technology.md

Latest commit

 

History

History
292 lines (199 loc) · 12.1 KB

technology.md

File metadata and controls

292 lines (199 loc) · 12.1 KB

Technology Stack

Seldon Vault is built on a modern, scalable architecture designed for reliability, cost efficiency, and transparency.

Multi-Provider LLM Factory

Seldon Vault integrates with 4 LLM providers:

Provider Models Used For
OpenAI GPT series Skeptic (deep review, medium reasoning)
Anthropic Claude series Seldon Arbiter (Opus with extended thinking + ReACT tools)
Google Gemini series Council debate member, fallback provider
DeepSeek DeepSeek series Analysts (Reasoner), Signal Processor & Translator (Chat)

Key design decisions:

  • Automatic failover with circuit breaker: tracks failures per provider, opens after 3 consecutive failures, resets after 5 minutes. Requests route to fallback provider seamlessly.
  • Retry logic with configurable timeout per call (default 300s). Bad request errors (400) skip retry and go straight to fallback.
  • Cost optimization: DeepSeek Reasoner for analysts ($0.02/call), DeepSeek Chat for preprocessing and translation ($0.001/call), GPT for skeptic review, Claude Opus for Seldon synthesis where reasoning quality matters most.
  • Provider switching requires no code changes — all LLM configuration centralized in seldon.yaml, not in agent code.

Data Pipeline

Continuous Signal Collection

12 data collectors run 4 times per day (not just once), pulling from diverse sources:

Collector Description
RSS Feeds ~63 feeds: Reuters, BBC, Al Jazeera, SCMP, Guardian, NYT, CNBC, Foreign Policy, and more
GDELT Global Database of Events, Language, and Tone
ACLED Armed Conflict Location & Event Data Project
FRED Federal Reserve Economic Data
Metaculus API Community forecasting platform
Polymarket API Real-money prediction markets
GDACS Global Disaster Alert and Coordination System
Fear & Greed Index CNN market sentiment indicator
Telegram 34 channel monitoring via Telethon
Reddit API 6 subreddits: geopolitics, economics, technology
Bluesky AT Protocol social feed

Incremental fetching: RSS uses HTTP conditional requests (ETag/If-Modified-Since) + Redis seen-URL cache (72h TTL). URL normalization strips tracking params before hashing. All collectors check Redis before returning posts.

Nightly reconciliation at 03:00 UTC: full batch reclustering using saved embeddings (zero API calls) corrects incremental drift.

Structural Data (Monthly)

Source Data
World Bank API Development indicators for 15 priority countries
IMF WEO World Economic Outlook projections
UN Population Division Demographic data and projections
OWID/GISS Climate data (temperature anomalies, emissions)

Dual-Persona System & Merge Layer

Three key domains use opposing Hawk/Dove persona pairs for cognitive diversity. The Merge Layer matches pairs by title similarity and produces enriched proposals with:

  • Classical weighted average probability
  • Disagreement spread metric
  • Quantum persona interference — wave superposition modeling: P = α²P_hawk + β²P_dove + 2αβ·cos(φ)·√(P_hawk·P_dove), where coherence reflects agreement despite opposite biases

Shadow mode: classical merge always determines actual probability; quantum result stored as metadata.

ReACT Tool System

The Seldon Arbiter uses iterative ReACT reasoning (Thought-Action-Observation loops) with 6 tools for the daily pipeline:

Tool Source What It Does
search_analogies Structural RAG (pgvector) Find historical parallels
query_indicators FRED API Fetch real-time economic data
fact_check Tavily web search Verify claims against live web
get_event_chain Event chain repository Explore temporal context
get_agent_track_record Agent calibration (Brier) Check analyst accuracy
compare_proposals Pure logic Analytically compare proposals

The Structural Seldon adds 2 more: compare_domain_briefs and get_previous_report for inter-epoch continuity. An additional ReACT tool analyze_cross_domain_dependencies finds entity/indicator overlap across domains.

Powered by Anthropic's Claude Opus with extended thinking — Claude reasons between each tool call.

RAG (Retrieval-Augmented Generation)

Two RAG systems serve different time horizons:

Daily RAG

  • pgvector embeddings (1536d, text-embedding-3-small) for historical analogy matching
  • Sector-filtered search with broadening fallback
  • Similarity threshold 0.65, top-3 results per domain

Structural RAG

  • Separate analogy bank for long-term patterns: Kondratiev waves, hegemonic transitions, debt crises, demographic transitions
  • Top-3 per domain + top-5 global analogies
  • Pattern types: kondratiev, debt_crisis, hegemonic_transition, technology_paradigm, demographic_shift, climate_regime, military_revolution

Bayesian Inference Engine

  • Likelihood Ratio method: posterior = (LR × prior) / (LR × prior + (1 - prior))
  • Confidence dampening: high = full LR, medium = 70%, low = 40%
  • Neutral band: LR within [0.85, 1.15] = no update
  • Per-horizon shift constraints: short-term ±15%/day down to century ±0.1%/day
  • Full probability history stored for every forecast

Adversarial Fact-Checking

Two-tier Skeptic system:

  • Quick Skeptic: Fast pre-filter with 6 structural kill rules (no web search, DeepSeek)
  • Max Skeptic: Deep review powered by Tavily Search API with real-time web search (GPT)
  • Media bias detection: availability bias, selection bias, narrative momentum
  • Evidence-backed critique with sources

Heuristic Alert Engine

Six deterministic post-synthesis sanity checks:

  • Overconfidence — P > 90% or P < 10%
  • Mid-range anchoring — P near 50%
  • Analyst disagreement — spread > 30%
  • Skeptic red flags — large adjustment in caution zone
  • Temporal mismatch — description vs. horizon inconsistency
  • Stale prediction — describes an already-occurred event

All alerts are informational — no auto-correction.

Cascade Propagation Engine

Graph algorithms detect causal chains between forecasts:

  • Link types: causes, amplifies, enables, triggers
  • Strength calibration: 0.3–1.0 with dampening per hop (0.5)
  • Maximum cascade depth: 3 hops
  • Batched propagation: all updated forecasts contribute to a shared pending pool before applying

Quantum Cascade Interference

When 2+ shifts target the same forecast ("fork target"):

total = classical_sum + interference_term
interference = 2 × √(|s_A| × |s_B|) × cos(θ)
θ = π × (1 - coherence)

Coherence from 4 factors: sector_match (0.30), direction_match (0.35), temporal_match (0.15), entity_match (0.20). Shadow mode — classical always applies, quantum computed alongside.

Knowledge Graph Engine

Three layers of semantic structure:

  • Signal Clustering — embedding cosine similarity (threshold 0.72) with 6-gate incremental append protection. LLM-generated summaries for multi-signal clusters
  • Source Reliability Ratings — per-source, per-sector Brier-based accuracy + independence scores. Zero LLM cost. Composite: 0.7 × accuracy + 0.3 × independence
  • Event Chain Linking — cross-day centroid cosine similarity (threshold 0.78). 8 lifecycle stages with regression protection. Chain-based forecast deduplication

Density Matrix Engine

Meta-uncertainty tracking for event chains:

  • 2-4 competing interpretations per chain with weights summing to 1.0
  • Purity = Σ(weight²) — tracks decoherence (1.0 = certain, 1/N = max uncertainty)
  • Sector templates for heuristic generation (zero LLM)
  • Keyword Bayesian updates during continuous ingest (zero LLM)
  • Daily LLM recalibration at 07:00 UTC for chains with purity < 0.85
  • Injected into Seldon and Skeptic prompts as formatted blocks

Decision-Maker Behavioral Profiling

14 world leader profiles loaded from YAML configurations:

  • Leaders: Trump, Putin, Xi Jinping, Erdogan, Netanyahu, Musk, Sam Altman, Jensen Huang, Dario Amodei, Kallas, Merz, Macron, Kim Jong Un, Powell
  • Profile fields: BVI (Behavioral Volatility Index), reaction speed, impulsivity, risk tolerance, priorities, decision circle, escalation patterns, historical patterns
  • Matched to signals by region and sector, injected into analyst prompts

Resolution Engine

Automated forecast verification:

  • Structured adapters: FRED (economic data), Yahoo Finance (assets/commodities), exchange rates, World Bank
  • LLM Resolver: web search + AI analysis for qualitative events
  • Three check strategies: on_date, periodic, near_deadline
  • Confidence-gated: high = auto-resolve, medium = flag, low = skip
  • Crisis protection: Seldon Crises never auto-resolved
  • Safeguards: stale evidence gate, hard evidence gate, minimum evidence threshold

Translation Layer

  • All agents produce English-only output
  • Dedicated translator (DeepSeek Chat) converts all user-facing fields to Russian
  • 1 forecast = 1 LLM call (all fields in one request for context coherence)
  • Multiple forecasts translated in parallel (5 concurrent)
  • Also translates: narratives, post-mortems, structural report commentary

Real-Time Updates

Server-Sent Events (SSE) provide push-based updates:

  • Event types: forecast_created, forecast_updated, narrative_created, structural_report_created
  • Frontend auto-invalidates React Query caches on events
  • Toast notifications for real-time awareness

Prompt Management

A 3-tier loading system:

  1. Memory cache (TTL configurable) — fastest
  2. PostgreSQL — persistent, editable at runtime
  3. YAML files — version-controlled fallbacks

Prompts are hot-swappable: change in database, takes effect within minutes. Supports base_agent inheritance for dual-persona variants.

Frontend

  • React 19 + TypeScript + Vite 7 — lazy-loaded pages (11 routes), ~300KB initial bundle
  • Tailwind CSS 4 — sci-fi terminal theme with circuit grid background and CRT scanlines
  • D3.js visualizations: force-directed cascade graph, TopoJSON world map, hexagonal cross-domain graph
  • Recharts for probability history and calibration charts
  • react-i18next for bilingual support (EN/RU) with automatic language detection
  • TanStack React Query for data fetching with SSE-based cache invalidation
  • react-helmet-async for dynamic SEO meta tags
  • 32 components including: ForecastCard, CascadeGraph, WorldMap, MasterScenarioPanel, CriticalJuncturesTimeline, DomainScenarioTree, DensityMatrixPanel, CrossDomainGraph

GEO (Generative Engine Optimization)

  • Schema.org JSON-LD: FAQPage, Dataset, BreadcrumbList
  • Dynamic XML sitemap with all forecast, narrative, and structural report URLs
  • Atom feed for RSS readers and crawlers
  • Prerender middleware for AI bot SEO (GPTBot, ClaudeBot, Googlebot)
  • robots.txt with explicit AI bot permissions
  • llms.txt for LLM discoverability
  • hreflang tags for EN/RU language alternates

Infrastructure

Everything runs in Docker containers:

  • PostgreSQL 16 + pgvector — embeddings, JSONB, async via asyncpg
  • Redis 7 — caching, dedup, concurrency locks, seen-URL sets, SSE
  • FastAPI — async REST API with rate limiting (slowapi, 3-tier: 60/min, 20/min, 10/min)
  • Celery + Redis — 15+ scheduled tasks across 2 worker queues (main + ingest)
  • Celery Beat — cron scheduler (00:00–08:00 UTC daily cycle + 4x/day ingest)
  • React + Nginx — reverse proxy with bot detection and prerender routing
  • Sentry — error tracking (optional)
  • Health checks on all services for automatic restart
  • Memory limits and structured logging (JSON in production)

See Also

  • How It Works — the full forecasting pipeline
  • Agents — meet the 11 analysts and their specializations
  • API Guide — endpoints for accessing forecasts, metrics, and real-time updates
  • README — project overview