VorkLab

LLM workflow optimization / implementation

Reduce repeated context

Deduplicate retrieved passages and keep stable instructions at the start of the prompt.

Prompt content order

InstructionsStable instructions & schemas
KnowledgeRelevant knowledge-base passages
TicketTicket & user data

A stable prefix.
Less repeated context.

Overlapping passages removed. Changing ticket data comes last.

Sonnet input tokens read from cache82.1%

Remove duplicate context and stabilize the prefix

Overlapping knowledge-base chunks were deduplicated. Static instructions, schemas and tool descriptions moved to the beginning of the prompt. Ticket and user data moved to the end.

The share of Sonnet input tokens read from cache rose from 80.7% to 82.1%. Caching already covered most input tokens, so this was a modest change. Smaller context and fewer calls contributed to the overall reduction. The report does not provide a controlled breakdown of savings by intervention.

Reading the cache measure

The percentage describes Sonnet input tokens billed as cache reads, excluding output tokens. It is not a cache-hit rate across requests or across both models. The diagram describes prompt content order, not a change to API message roles.

Combined results and evaluation

See the shared financial calculation, rollout checks and scope of the reported results.

Read the evidence behind the case
Agentic Code-Level Cloud & LLM Optimization
Request a free billing review

Accepted projects have no upfront service fee. Fixed fee: 30% of verified net savings over 12 agreed months, after approved transition and additional operating costs.