TokenSaver
EN

La documentation est disponible en anglais.

Documentation / Use cases

Cache & savings

Turn on the Cache module for a key so repeated, equivalent prompts can return a hit and skip the upstream LLM — while still appearing in Flux IA.

Cache is the fastest win for demos and production bots that re-ask similar questions (onboarding scripts, eval harnesses, FAQ agents). Savings show up in Cost Management and on overview KPIs as cache hit rate.

Cache hit path
  1. Prompt
  2. Cache lookup
  3. Hit → skip LLM
  4. Flux IA still records the op

First call

Miss: pipeline may compress / RAG / PII, then call the provider. Tokens billed include the LLM.

Second identical call

Hit: response served from cache. Provider not called. Dashboard shows hit; graph still has a hub.

  1. Enable Cache policy Governance → API key → Cache ON. Confirm the Modules panel shows Cache actif for that key.
  2. Send the same prompt twice Use SDK ask(), curl on /openai/v1, or any Connect tool. Keep model and message content identical.
  3. Check Flux IA + costs Second row should indicate cache. Overview / Cost Management should reflect lower billed tokens.
Flux IA — cache hit row
Cost / overview — cache KPI
# 1) Governance: Cache ON for ts_…
# 2) Call twice with the same body
curl -sS "$OPENAI_BASE_URL/chat/completions" \
  -H "Authorization: Bearer ts_…" -H "Content-Type: application/json" \
  -d '{"model":"gpt-4o-mini","messages":[{"role":"user","content":"Ping"}]}'
# Repeat immediately — expect cache hit in Flux IA