How We Translate Entire Books with LLMs Without Losing Context

Short summary

LectuLibre translates books using LLMs by splitting chapters into sentence-based chunks (1800 tokens max) rather than character breaks, preserving narrative flow. To maintain coherence across chunks without doubling token costs, they use a 5-sentence sliding context window from the previous chunk plus a one-sentence running summary fed into every prompt. This reduces token usage by ~80% compared to including full previous chunks.

• •Sentence-based chunking preserves narrative coherence vs character/paragraph splitting
• •Sliding context (5-10 previous sentences) + running summaries provide long-range coherence without doubling token costs
• •1800-token chunks optimize API call speed and cost while maintaining translation quality

Generated with AI, which can make mistakes.