A retriever searches an external document index and returns passages that might answer a query. In retrieval-augmented generation (RAG), it supplies evidence to a reader or generator rather than writing the answer itself. If it misses the useful passage—or ranks a lookalike above it—the downstream model may have little chance to answer well. section 1.6 section 3.0
Choose the retrieval objective before choosing the model:
- For exact terminology and strong query–document vocabulary overlap, start with keyword retrieval such as BM25. It is fast, inexpensive, and avoids some semantic-search ambiguity. section 2.1.6
- For conversational queries that use different wording from the documents, consider semantic retrieval. A dense encoder maps queries and passages to vectors for similarity search; question-answering benefits from learning which passages answer questions, not merely which sound similar. section 2.1.6 section 3.0
- Match ranking to the reader. A reader handling few passages needs high precision near the top; Fusion-in-Decoder can work with a recall-oriented top-k set, but extra passages add noise and cost. Measure the useful range rather than assuming more is better. section 5.3.1
Table 3.1 Symmetric search vs asymmetric search
|
Feature
|
Symmetric
Search
|
Asymmetric
Search
|
|
Goal
|
Find similar text
(twins)
|
Find answers to
questions
|
|
Input
|
Similar
length/structure
|
Short query vs.
Long doc
|
|
Model
|
SBERT
|
DPR
|
|
Use Case
|
Paraphrase
detection
|
Open Domain QA
|
The comparison distinguishes symmetric similarity search from asymmetric question-to-answer-passage retrieval—the distinction that matters when a short question must find evidence in a longer document. section 3.1
Indexing and querying are separate decisions. The document encoder produces vectors for the searchable corpus; at query time, the query encoder produces a vector and the retriever ranks candidates. In canonical RAG, DPR’s document encoder and prebuilt index stay fixed while its query encoder is fine-tuned with the BART generator. Keeping the index external also allows knowledge to be refreshed by changing indexed documents rather than retraining the generator, though retrieval adds an inference step. section 4.1 section 6.1
A ranking failure in financial search
An analyst asks why operating margins declined. A sentence repeating the company, quarter, and margin figures ranks above the passage explaining the cause. Both are indexed, so widening top-k exposes the answer, but does not fix the ordering. This is a reason to evaluate answer-bearing rank—not just whether the document exists in the corpus. section 6.7
Training can target that failure directly. REALM jointly trains a retriever and language model so useful evidence improves the language-model objective; its retriever still performs vector search, but training changes the encoder weights and thus the ranking. Atlas likewise uses reader performance to favor useful documents. These approaches offer stronger alignment than a static similarity retriever, at greater training complexity; a pretrained, modular retriever remains a practical baseline. section 3.0 section 3.2 section 6.8





