All documents

Interactive

Break the assistant.

This is the AI side of the Service Assistant, wired up and actually running. Turn the traffic up past a million requests a day, starve the cache, kill a graphics card mid answer. Poke it however you like and watch what the design does about it. The interesting number is not how fast one card is. It is how few requests ever reach one.


Fig. 1 · our AI, under your control Live
cache retrieval gate queue gpu cards out fallback requests api cache · retrieval queue gpu cards out
steady 3 api 2 cards queue 0 batch 6

Api tier

where a request lands first, and the only part that scales by adding cheap machines

Cache

a repeat question, answered without touching anything downstream

Retrieval

finds the passage in that community's documents; arithmetic, not a model

Gate

nothing close enough, so it is refused here and no card is ever asked

Queue and cards

what is left is batched; one card answers many questions per pass

Fallback

drawn alongside, taken whenever the cards cannot

Served
0/s
That is per day
0
Never reach a card
0%
Queue
0
Response
0ms
Job wait
0ms
Cost / 1k
$0.00

    ControlsRunning

    Traffic

    A million requests a day is about twelve a second on average, and three times that at the morning peak. Drag it past either.

    arrival rate12 req/s
    burstiness2.0x

    Cache

    Residents ask the same twenty questions. A repeat is answered from memory and never reaches retrieval, a queue, or a card.

    cache enabled
    hit rate60%

    The gate

    A question the documents cannot answer is refused here, with no model called at all. That is honesty first, and capacity second.

    refused here18%

    Gpu cards

    A card does not answer one question at a time. It answers a batch each pass, so throughput rises with batch size and so does the wait. Autoscale is on: a burst buys a card and then gives it back.

    cards2
    batch size6
    autoscale
    ceiling6 cards

    Api tier

    Cheap machines. They do no thinking, so more of them costs almost nothing and buys headroom at the front.

    instances3

    Break it

    A slow dependency is worse than a dead one. The breaker turns slow into fast and degraded, on purpose.

    A model, not a measurement. Retrieval, the gate and the fallback behave as the running system does, and the refusal rate is the one measured on real questions. The card figures are modelled on an NVIDIA L4 running an 8B model with continuous batching, where a pass costs roughly 900ms plus 115ms per item in the batch. Nothing on this page connects to the live service.