Query-document dual encoders

An encoder contract includes tokenizer, truncation, prompt or prefix, pooling, model weights, output dimension, dtype, and normalization. Query and document encoders may share parameters or be distinct, but their outputs must be trained to interact through the deployed similarity function. A model trained for unnormalized inner product can use document magnitude as part of its score. Normalizing those vectors afterward can therefore change rankings rather than merely improve numerical stability. A new checkpoint can require re-encoding the corpus rather than reusing vectors with the same shape.

Cosine similarity divides by both nonzero vector norms. On unit-normalized vectors, cosine and inner product coincide, and squared Euclidean distance becomes a monotone inverse transformation of that product. Without those conditions, the rankings need not agree. Faiss’s official metric documentation distinguishes maximum inner product from squared L2 search and describes normalization for cosine. Check whether an API expects larger similarities or smaller distances; a sign or ordering mistake can return the least similar candidates while still producing valid-looking numbers.

First compare a small exact search over the intended vectors against the production search path. If the exact ranking already misses the relevant document, changing approximate-index settings cannot repair that representation failure. If exact search succeeds but approximate search does not, inspect index construction, metric, search effort, quantization, and version compatibility. Both effects can coexist. Keep a stable vector snapshot and exact reference for debugging instead of attributing every dense retrieval miss to semantic model quality.

Training and deployment also need aligned document units. A passage encoder trained on bounded chunks may not preserve relevance when fed a much longer document that truncates the answer. Query prefixes or language handling can change the embedding distribution. Evaluate identifiers, negation, dates, multilingual queries, and domain-specific vocabulary with appropriate judgments. DPR’s experiments demonstrate a useful method on specific tasks, not a theorem that dense retrieval replaces every lexical signal or succeeds on every new collection.

Publish index metadata that identifies the document encoder and processing pipeline. Ensure the serving query encoder belongs to the intended paired version, and make reindex or migration behavior explicit. A partial rollout that mixes incompatible vector versions can create class-dependent failures difficult to see in averages. Compare exact retrieval, candidate recall, final reranking quality, and latency separately. A successful vector-dimension check catches only one narrow class of mismatch and does not establish that the embeddings inhabit the same learned space.

The vector lab illustrates two-dimensional geometry only. Reset, set angle to 0 degrees and magnitude to 1, predict the horizontal component, then Step: one. Compare a fresh Reset with both angle at 60 degrees and magnitude at 1 before Step: one half. Against a fixed unit query on the horizontal axis, that component equals the inner product. The instrument does not run an encoder, tokenize documents, build an approximate index, or measure semantic relevance.