Live product
Hannim
Result
40
LLM routes over one pgvector corpus
Problem
TOPIK II learners get either a graded reader or an AI tutor, never both grounded in the same corpus, so feedback contradicts the material.
What broke
Embedding the corpus hit Cohere's 100k tokens/minute throttle mid-ingest and failed partway, leaving the store inconsistent.
Batching the work off-peak and making ingest resumable fixed both the throttle and the partial-state problem.
Approach
Forty API routes over a single pgvector store — reader, mock exams, tutor, listening drills, conjugation practice and a mistakes bank all retrieve against the same embedded corpus.
Stack Nuxt · pgvector · OpenRouter · Groq · Cohere · PostgreSQL
Next project →Law4Kor