Lexical / Keyword Retrieval
Our retrieval stack preserves lexical search because exact language remains a first-class retrieval requirement.
Where exact terms decide
Embeddings are useful for semantic similarity. They are not a replacement for exact terms. Lexical retrieval is often the correct path for:
- Customer or employee names
- Account numbers
- Contract language
- Product codes
- Dates
- Quoted phrases
- Statutory or policy wording
- Unique identifiers
Exactness as a feature
Consider the difference between searching for documents about late-payment disputes and searching for the exact contract phrase “payment due within 30 days.”
The first is semantic. The second depends on the words themselves.
Our retrieval architecture can therefore maintain lexical indexes alongside semantic indexes and structured data access.
- Yes → lexical candidate generation
- No or mixed → hybrid route
Tokenization, normalization, field weighting, phrase search, filters, and metadata can all affect the resulting candidate set.
Where lexical retrieval fails
Exact search can miss relevant records when the wording changes. A customer may describe the same complaint using several phrases. Employees may use abbreviations, shorthand, or different terminology. Historical records can use names that later changed.
That is why lexical retrieval is not treated as the universal answer either. Entity resolution can connect aliases. Semantic retrieval can expand conceptual recall. Structured filters can constrain the correct customer or period. Reranking can order mixed candidate sets.
Precision on exact terms remains valuable even in an AI-native system. We do not discard deterministic search simply because a probabilistic representation is available.
Lexical search inside hybrid retrieval
Lexical scoring can be combined with field-aware weighting and metadata constraints so that an exact match in a contract identifier does not have the same meaning as the same token appearing incidentally in a long transcript. The index can also preserve phrase and field semantics that are difficult to recover reliably from vector similarity alone.
This makes lexical retrieval especially useful as a precision channel inside hybrid search. It can generate exact candidates, verify semantic candidates, or act as the primary path when the query itself indicates that exact wording matters.
The broader principle is straightforward: AI-native retrieval should add useful search modes without discarding the deterministic ones that remain better for exact evidence.