Open-domain question answering answers questions across nearly any topic by drawing on a massive external corpus rather than a subject-specific collection. It differs from open-book tasks: open-domain describes the breadth of possible questions, while open-book describes permission to consult sources such as Wikipedia or an enterprise database. section 3.1
A standard late-2010s architecture separated fast retrieval from deep reading. Sparse retrievers such as TF-IDF and BM25 searched millions of documents and selected likely passages; a reader such as BERT then extracted an answer. Later systems used dense retrieval and generation, trading exact grounding against synthesis and computational cost. section 3.1 section 4.5
Common approaches differ in where knowledge resides and what the system returns:
- Parametric models such as T5 answer from memorized knowledge, but knowledge freezes at training time and hallucination remains possible. section 4.5
- DPR retrieves relevant passages and extracts exact answer spans, grounding factual lookups but struggling when answers require synthesis. section 4.5
- RAG combines a frozen DPR retriever with a BART generator, joining document grounding with generated answers through modular fine-tuning. section 4.5
Wikipedia-scale question answering
The Natural Questions benchmark asks systems to answer authentic questions such as “What is the ozone layer and how does it affect Earth?”, “Who plays Voldemort in the Harry Potter movies?”, and “When did the Soviet Union collapse?” from 21 million Wikipedia passages. section 4.5
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
|
Use the comparison to distinguish symmetric similarity search from asymmetric question-to-document retrieval, the pattern required for open-domain question answering. section 3.1

