Hybrid Retrieval
Our retrieval architecture combines multiple retrieval methods because company knowledge does not have one universal search surface.
One index is not enough
A semantic query, exact contract clause, account identifier, date constraint, customer relationship, and structured financial field should not all be resolved through the same index. Hybrid retrieval allows the execution path to combine the signals appropriate to the task.
Retrieval channels
A candidate set can be assembled from semantic retrieval, lexical or keyword retrieval, metadata filters, structured queries, entity constraints, relationship constraints, and temporal filters. The resulting candidates can then be reranked and assembled into task-specific context.
- Semantic
- Lexical
- Structured
- Entity / relationship
- Time / metadata
Why one method is insufficient
Semantic similarity is useful when the wording differs but the meaning is related. It is weaker when the task depends on an exact identifier, amount, quoted clause, or date.
Lexical retrieval excels at exact terms but can miss conceptually equivalent language. Structured queries are precise where a schema exists but cannot explain the reasoning contained in a call or email.
Hybrid retrieval preserves those differences rather than forcing every source into one representation.
Evaluation
The retrieval objective is not to maximize the number of documents returned. We evaluate whether the pipeline retrieves the evidence necessary for the actual task with acceptable precision, recall, latency, and context cost.
More retrieved material can reduce answer quality if irrelevant records displace or obscure the correct evidence. Our retrieval layer therefore treats candidate generation, reranking, filtering, and context construction as separate decisions.
The correct retrieval path is selected by the workload, not by allegiance to a particular search technique.
Routing before ranking
Hybrid retrieval does not require every query to hit every index. The task can first determine which retrieval channels are plausible, reducing unnecessary candidate generation and latency.
An exact account identifier may start lexically or structurally; a conceptual operating question may start semantically; a point-in-time investigation may apply entity and temporal constraints before either. Reranking then operates on the candidate set rather than replacing the initial retrieval methods.
This separation is important because candidate generation optimizes recall while reranking can optimize relevance among the retrieved candidates. The final context can still be smaller than the candidate pool, preserving model attention for the evidence most likely to matter.