Search, Ask, Chat: Three Kinds of Retrieval

Open Notebook does not have one universal way to retrieve research. It offers lexical text search, semantic vector search, the multi-stage Ask workflow, and conversational Chat over selected context. This chapter compares those paths by the question each one answers, the material it can see, and the kind of mistake it can make. The practical value is conditional choice: use exact retrieval to verify words and numbers, semantic retrieval to discover related ideas, Ask to assemble a broad response, and Chat when you know which material deserves a continuing conversation.

Four ways to decide what matters

The first distinction is between finding a phrase and finding an idea.

Text search is lexical. It looks for words and phrases in the material held by the knowledge base. In the current search functions, that includes source titles and full text, source chunks, source insights, and note titles and content. Results are grouped and ranked by lexical relevance. If a researcher remembers that a report uses the phrase “delivery reliability,” text search is the direct tool. It can also find a number, proper name, technical term, or exact quotation that semantic similarity might treat as unimportant.

Lexical search is strongest when the words themselves are evidence. It is weaker when the researcher knows the concept but not the author's vocabulary. Searching for “late delivery risk” may miss a passage that discusses “supplier performance volatility.” The underlying idea may be present even when the exact terms are not.

Vector search addresses that second problem. It turns the query into an embedding and compares it with embeddings for source chunks, source insights, and note content. The result is ranked by semantic similarity. A question about “how suppliers fail to meet deadlines” may find a passage about delivery volatility even if it does not contain the word “risk.” Vector search is therefore a discovery mechanism: it helps the researcher find candidates whose meaning seems related.

That strength is also its boundary. A vector result is not an exact quotation, and source or note titles are not matched semantically in the current search path. Rare identifiers, precise numbers, and technical terms may be better served by text search. Similarity can also connect ideas that are related in topic but different in claim. The higher-ranked passage is a promising place to look, not a verdict about what the source proves.

Notebook scope adds a separate decision. Search and Ask can operate over the wider knowledge base when no notebook is selected, or over the sources and notes connected to one or more chosen notebooks. Scope is resolved through the notebook relationships rather than inferred from a title. A source linked to any selected notebook can participate; a note must be linked through its notebook relationship. This lets a researcher keep public background material and project-specific material in different working spaces, but it also means a search result can change when the scope changes.

The second distinction is between retrieval and synthesis. Search returns candidates. Ask goes further. It receives one broad question, asks a strategy model to formulate several search terms, runs a vector search for each term, asks another model to interpret the results for each search, and sends those intermediate answers to a final model for synthesis. The current graph allows up to five strategy searches and retrieves up to ten results for each fan-out.