When a question-answering system needs current information, keeping documents in an external index is often more practical than retraining its language model. Contriever is a dense retriever that represents text as vectors. It learns from unsupervised contrastive training rather than human-labeled query–document pairs, making it useful when relevance labels are unavailable. section 6.1 section 7.3
In Atlas, Contriever supplies the retrieval component. Transformer-based encoders embed queries and documents; document vectors are computed ahead of time and stored in an index. At query time, dot-product similarity ranks documents against the query vector. The retrieved information can change when the index is updated, without changing the language model’s parameters. section 6.1
Searching with a hypothetical answer
HyDE takes a different route to the query vector: it generates a hypothetical answer passage, then embeds that passage with the same encoder used for the real documents. Contriever was the encoder in the original HyDE work. Its fixed-size embedding can discard an incorrect date or name while retaining the passage’s topic. It cannot rescue a passage framed around the wrong subject: retrieval may follow that mistaken framing. section 7.3
Table 7.1 HyDE zero-shot performance vs. baselines
|
Benchmark
|
BM25 (Lexical)
|
Contriever (Unsupervised Dense)
|
HyDE (Zero-Shot)
|
Contriever-FT (Supervised Dense)
|
|
TREC DL19
|
38.3
|
44.5
|
61.9
|
64.5
|
|
TREC DL20
|
35.8
|
43.1
|
60.1
|
62.1
|
|
BEIR (SciFact)
|
66.8
|
64.3
|
69.7
|
71.0
|
|
BEIR (FiQA)
|
26.9
|
30.5
|
38.8
|
48.1
|
The benchmark table places unsupervised Contriever beside lexical BM25, zero-shot HyDE, and supervised fine-tuned Contriever. It shows why adding a hypothetical passage can help, while giving a reason to evaluate retrieval on the intended dataset rather than assume the more elaborate pipeline always wins. section 7.3
