--- ## Problem/Motivation `ai_translate` translates **one field-column per chat request**. `TextTranslator::translateContent()` is called once per column, and `ChatTranslationProvider::translateText()` rebuilds the full prompt (the configured translator prompt template + `setSystemPrompt('You are a helpful translator.')`) and calls `chat()` for each one. For a node with N translatable field-columns, that is N requests, each re-sending the **same** instruction block. For a typical 3-field node, roughly 1,000 of \~1,500 input tokens are the identical instructions sent three times. The redundancy scales linearly with field count and node count. **Prompt caching does not solve this.** The translator prompt is only \~500 tokens, well below the minimum cacheable prefix of common models (e.g. 4,096 tokens on Anthropic Haiku 4.5), so the prefix silently never caches (`cache_creation_input_tokens` stays 0). The fix is to send the fields together. On a large backfill the cost is real: across \~35,000+ nodes this is on the order of tens of dollars of avoidable input cost, plus \~Nx the request count — which also increases latency and rate-limit pressure. ## Proposed resolution Add an **opt-in, additive** batch path that packs all of a node's translatable field-columns into a **single** chat request, leaving the existing per-field path completely unchanged. - New method `translateMultiple(array $texts, LanguageInterface $langTo, ?LanguageInterface $langFrom, array $context): array` on the ai_translate text translator (`TextTranslator` + `TextTranslatorInterface`), keyed by caller-supplied field id. It operates against the **`chat`** operation type directly (mirroring what `ChatTranslationProvider` already does internally), since batching many texts into one prompt is a chat concern — native `translate_text` providers (DeepL/Google-style) are single-text by design. - `translateContent()` (the single-field path) is **unchanged**. When no usable chat provider is configured, `translateMultiple()` transparently loops `translateContent()` per field — so enabling the feature never changes behavior or adds a hard requirement. - **Wire format: a robust delimiter, not structured JSON.** The `ai` module exposes no capability flag for structured output (the `AiProviderCapability` enum has only `StreamChatOutput` / `ChatFiberSupport`), so structured-output support cannot be feature-detected, and only OpenAI-based providers honor `ChatInput::setChatStructuredJsonSchema()`. JSON would also force every field's HTML through an escaping round-trip. Instead, each field is wrapped between unique sentinel markers (field id + a per-request random nonce); HTML rides raw between markers and is parsed back by id. This works on **every** chat provider. - **Partial failure degrades gracefully:** a whole-request failure falls back to per-field for every field; a single field that doesn't round-trip (missing / duplicated / empty / broken sentinel boundary) falls back to per-field for just that field. No request is retried more than once. Each fallback is logged. - **Large nodes are chunked** by a configurable character-count threshold (provider-agnostic; token counting isn't portable across `ai` providers). Fields are greedily bin-packed into batches under the threshold; a single oversized field is sent on its own. This also mitigates the large-content timeouts reported in #3545381. - **Opt-in:** the batch path is exposed as a separate method (and/or config flag) so no consumer behavior changes until a caller opts in. ## Remaining tasks - [ ] Implement `translateMultiple()` + interface addition (additive; per-field path untouched) - [ ] Sentinel build/parse with per-request nonce; per-field fallback on any field that doesn't round-trip - [ ] Character-based chunking / greedy bin-packing; oversized-field handling - [ ] Carry the existing prompt-injection guard into the batch prompt + add data/instruction framing - [ ] Tests: multi-field round-trip, HTML preservation, sentinel parse + per-field fallback, partial-failure behavior, oversized-field chunking - [ ] Demonstrate token reduction with before/after token counts on a representative multi-field node - [ ] phpcs (Drupal + DrupalPractice) clean; GitLab CI green - [ ] Maintainer review of API placement and opt-in surface ## User interface changes None by default. Optionally a config setting for the batch-chunk character threshold and/or a toggle to enable batching. No change to existing behavior unless enabled. ## API changes / Data model changes - **Additive only.** New `translateMultiple()` on `TextTranslatorInterface` (a default implementation keeps the interface non-breaking for existing implementers). No change to `translateContent()` or to the `ai` module's `TranslateText` operation type. ---