How AI OS Works

Five external data sources feed an enrichment pipeline. The processed stream passes through an LLM router and a vector RAG store before reaching three consumers: the public portal, the admin dashboard, and the MCP server.

Each task is first classified by type, then dispatched to the backend that maximises quality per unit cost and latency. Fast drafts go to Groq; structured extraction and reasoning to OpenAI; bulk jobs to Runpod; embeddings and free-tier work to Ollama.

Task → backend dispatch table

Cost-optimisation formula

cost(task, backend)    = price_per_token × estimated_tokens(task)
latency(task, backend) = base_latency + tokens / throughput
quality(task, backend) = exam_score(backend, task_type)

route(task) = argmax( quality / (cost × √latency) )

Every enrichment cron is benchmarked against its expected interval and a hard SLA. Health degrades through three zones as elapsed time grows — OK → degraded → critical.

health(cron) =
  1   if last_run_age < 1.5 × expected_interval   → OK
  0.5 if last_run_age < SLA                        → degraded
  0   otherwise                                     → critical

Agent performance is measured on five axes and collapsed to a single effectiveness score E via a complexity-weighted mean. Heavier tasks contribute proportionally more to the final score.

E = Σ(wᵢ × scoreᵢ) / Σ(wᵢ)

where:
  wᵢ      = complexity weight of task i
  scoreᵢ  = outcome score (0 – 100)

Axes (5 dimensions):
  Speed    — task latency vs budget
  Cost     — tokens × price per token
  Quality  — exam score on task type
  Safety   — guardrail pass rate
  Learning — hint reuse in context pack

Each query is embedded into a 768-dimensional vector, matched against LanceDB chunks by cosine similarity, re-ranked by recency decay, and packed into the model's context window.

query → embed(query) → cosine_similarity(query_vec, chunk_vecs)
      → top-K chunks → rerank by recency
      → context_pack → inject into prompt

GOGA classifies the current market regime from real-time signals and selects the appropriate execution strategy. Each row maps a detected condition to its indicator set and the action taken.