ADR-005: TYPO3 Caching Framework Integration 

Context 

LLM API calls are:

• Expensive (cost per token).
• Relatively slow (network latency).
• Often deterministic (embeddings, some completions).

Decision 

Integrate with TYPO3's caching framework:

• Cache identifier: nrllm_responses.
• No backend specified — TYPO3 uses the instance's default cache backend (respects Redis/Valkey/Memcached).
• Cache keys based on: provider + model + input hash.
• TTL: 3600s default (configurable).
• Cache group: nrllm (flush via cache:flush --group=nrllm).

Caching strategy:

• Always cache: Embeddings (deterministic).
• Optional cache: Completions with temperature=0.
• Never cache: Streaming, tool calls, high temperature.

Consequences 

Positive:

• ●● Reduced API costs.
• ●● Faster responses for cached content.
• ● Follows TYPO3 patterns.
• ◐ Configurable per deployment.

Negative:

• ✕ Cache invalidation complexity.
• ◑ Storage requirements.
• ✕ Stale responses if TTL too long.

Net Score: +4.5 (Positive impact - significant cost/performance gains with manageable cache complexity)