fix: enforce embedding model max token limit via proper truncation

[timon0305]

Summary

Fixes the embedding model max token limit not being enforced, causing API errors when input text exceeds the model's context window. This was reported across multiple providers (Ollama, vLLM, HuggingFace TEI) with models like bge-large (512 tokens), multilingual-e5-large-instruct (512 tokens), etc.

Root causes fixed

1. LLMBundle.encode() truncated by character count instead of token count — text[:target_len] slices by character index, but max_length is in tokens. For multi-byte text (Chinese/Japanese), this still exceeds the token limit. Fixed to use truncate() which properly tokenizes and truncates at the token level.

2. LLMBundle.encode_queries() had no truncation at all — User search queries passed directly to embedding API without any length check. Added token-level truncation using the configured max_length.

3. 12 embedding provider classes lacked input truncation — Added truncate() calls to both encode() and encode_queries() for: OllamaEmbed, XinferenceEmbed, LocalAIEmbed, HuggingFaceEmbed, CoHereEmbed, SILICONFLOWEmbed, ReplicateEmbed, BaiduYiyanEmbed, VoyageEmbed, VolcEngineEmbed, ZhipuEmbed (encode_queries), QWenEmbed (encode_queries character slice → token truncate).

4. Misleading tooltip for max_tokens setting — The UI tooltip said "Defaults to 512" but the actual default is 8192. Updated en/zh locales to accurately describe the field as "maximum input context length" with correct default.

Changes

• api/db/services/llm_service.py — Fixed LLMBundle.encode() to use token-level truncate() instead of character slicing; added truncation to LLMBundle.encode_queries()
• rag/llm/embedding_model.py — Added truncate() to 12 provider classes missing input truncation
• web/src/locales/en.ts — Fixed maxTokensTip tooltip text and default value
• web/src/locales/zh.ts — Fixed maxTokensTip tooltip text and default value (Chinese)

Closes #4683

[dosubot]

Related Documentation

Checked 0 published document(s) in 1 knowledge base(s). No updates required.

^{How did I do? Any feedback?}