How to search
RBAi Scholar.

1. The three modes

2. Which one do I want?

A useful rule of thumb: if an empty result page would surprise you, use keyword mode. Keyword mode tells you plainly that nothing matched your terms. Semantic mode always returns its closest guesses, which is what you want when exploring and not what you want when verifying.

3. Where Boolean operators work

This is the part worth reading twice, because the two modes differ and the difference is not guessable.

The trap to avoid

In semantic or hybrid mode, this does not do what it looks like:

"TP53" OR "BRCA1" in breast cancer

Because OR is not an operator here, both quoted strings are treated as requirements. The search returns records containing TP53 and BRCA1 together — the opposite of what you asked for — and the word OR is encoded as part of your sentence. The API now returns a notice saying so, but the simpler fix is to switch modes:

{ "mode": "keyword", "query": "TP53 OR BRCA1" }

4. Quoted phrases, and what adjacency means

A quoted phrase of several words requires those words to appear next to each other, in that order — not merely somewhere in the same record.

For "alpha synuclein aggregation":

This is why quoting a phrase is far more selective than quoting each word separately. "Safe Return" would be bounded at roughly 140,000 records by its commonest word, but as an adjacent phrase it matches 131.

Quoting is also how you defend a string the language model would otherwise dissolve. A rare identifier such as R175H or PD173074 is split into fragments by the encoder and averaged with every other word in the record, so the one string that makes a paper unique gets diluted. Quoting it makes it a requirement instead.

5. Hybrid search: a question with a required string

Write a normal question and put quotation marks around the part that must appear:

role of "MDM2" in stabilising mutant p53

Two things then happen at once, and the second surprises people:

1. The quoted string becomes a hard requirement. Records that do not contain MDM2 are not considered at all.
2. The quoted words are still part of the meaning search. Only the quotation marks are removed before your sentence is encoded. The words stay exactly where you put them.

So the text encoded for the example above is:

role of MDM2 in stabilising mutant p53

The marked string therefore does double duty: it narrows the candidates and contributes to the meaning. Quoting a term does not remove it from the question.

Where you put the quoted term in the sentence makes no measurable difference to the results, because the requirement is evaluated without reference to position.

6. Why a search is sometimes refused

Two guards will decline a query rather than answer it slowly. Both return a clear message with the measured number, not a generic failure.

A quoted string that is too common

Requiring a string that appears in more than 20,000 records does not narrow the search usefully, and it makes the search slow. Requiring "ARS", which is in about 38,000 records, is declined with that figure and a suggestion.

What to do: quote something rarer, or quote a longer phrase. A multi-word phrase is much more selective than its commonest word, so "aminoacyl-tRNA synthetase" succeeds where "ARS" is refused.

A keyword expression that matches too much

Ranking cost grows with the size of the match set, so a keyword expression estimated to match more than 25,000 records is declined. Add a term, use AND instead of OR, or quote a phrase.

7. Worked examples

Two things not to do

• Do not send a full sentence to keyword mode. Bare terms are combined with AND, so a twenty-word question demands that one record contain all twenty words. Almost nothing will.
• Do not send bare keywords to semantic mode. The encoder is built for sentences; three disconnected nouns give it very little meaning to work with.

8. What the API tells you about your query

Every search reports how it interpreted what you sent, so you never have to guess. A semantic response includes:

{
  "mode": "semantic",
  "semantic_query": {
    "hybrid": true,
    "encoded": "role of MDM2 in stabilising mutant p53",
    "required_phrases": ["MDM2"],
    "tsquery": "'MDM2'"
  },
  "notices": [
    { "code": "hybrid_search_active",
      "message": "Records must contain \"MDM2\". The rest of the question is matched on meaning." }
  ]
}

encoded is the exact text that went to the language model. required_phrases is what the records had to contain. Keyword mode reports the parsed expression in keyword_query in the same way.

notices is advisory and never fatal. The one worth watching for is boolean_operators_not_parsed_in_semantic_mode, which means you typed AND, OR or NOT somewhere it has no effect.

9. Known limits

Stated here rather than left to be discovered:

• No parentheses. Grouping is unsupported in both modes. a OR b AND c is evaluated left to right under PostgreSQL's precedence, which may not be what you intended. Split the search into two queries instead.
• Boolean operators do nothing in semantic and hybrid mode. They are encoded as ordinary words. Use keyword mode.
• A quoted phrase containing a common short word such as of or the may behave unexpectedly, because those words are dropped from the index before adjacency is checked.
• Case is ignored for matching, apart from the uppercase operator keywords in keyword mode.
• Semantic mode never returns nothing. It returns its closest guesses. Use keyword mode when you need to know that no record matched.

Related

• RBAi Scholar — the search interface, coverage and API access.
• RBAi MCP Cloud — using the corpus from Claude, Cursor or VS Code.