GenixBit Genius
Architecture Whitepaper

Latency Exponential Moving Average (EMA)

The AI Routing layer utilizes Redis to track model latencies dynamically. Upon every successful API call, we calculate a moving average response time using an exponential decay filter ($\alpha = 0.2$):

If a model endpoint throws a rate-limit error, timeout, or HTTP 5xx exception, the router applies a temporary 10.0-second latency penalty. This automatically routes subsequent user queries away from the offline provider during failover windows.

Failover Traversal Queue

Each model ID contains a fallback chain mapping to alternative models in similar capability classes:

• gpt-4o → gemini-1.5-pro → claude-3-5-sonnet → deepseek-v3
• llama-3-70b → qwen-2.5-72b → mixtral-8x7b

When a query is dispatched, the backend traverses this queue sequentially. If a provider call fails, the transaction is immediately caught and retried on the next provider inside the current task execution context, avoiding client-side error propagation.

Vector Retrieval Mechanics

Semantic searches are executed inside PostgreSQL using the pgvector cosine distance operator (<=>). Chunks are sorted by distance and filtered by user access keys:

SELECT c.content, c.chunk_index, d.filename,
       (c.embedding <=> :query_embedding) as distance
FROM document_chunks c
JOIN documents d ON c.document_id = d.id
WHERE d.user_id = :user_id
ORDER BY distance ASC
LIMIT :limit