Portfolio-grounded Liquid answers

Every Jekyll build emits /assets/data/portfolio-search.json from published posts, publications, teaching materials and talks. Set search_exclude: true to omit an item. Drafts, private databases, visitor conversations and voice recordings are excluded. For a PDF or upload, publish a reviewed text companion with its source link in a supported collection. Arbitrary uploads are not automatically ingested.

The browser chunks text and ranks it with BM25, passing up to three passages into the existing fine-tuned Liquid model. English words and Korean bigrams provide lexical matching, not cross-language semantic search. Searches stay local.

Free-form model answers first perform a host-directed search. The model can emit <search_portfolio>query</search_portfolio> for one additional read-only search; the host executes it and returns passages for model synthesis. This is a bounded text protocol, not native JSON tool-call training. Existing privacy, entity and verified-profile routes remain in place. Retrieved source links are appended by the host, not invented by the model.

The prompt separates explicit beliefs from interpretation and treats passages as data, not instructions. This is not a guarantee against hallucinations or prompt injection: curated public ingestion remains necessary. Sources identify retrieved material, not verified entailment of every generated claim.

Run node --test scripts/test_portfolio_search.mjs and build Jekyll. Before merging, evaluate the actual quantized checkpoint on held-out philosophy questions, follow-ups, no-evidence questions, malicious quotations and bilingual paraphrases. Measure recall@3, unsupported claims, abstention and citation relevance. New SFT weights, automatic PDF ingestion and native tool-call training are not included.