AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。
Learn Technical Articles Relevance Feedback in Qdrant Back to Search Quality Relevance Feedback in Qdrant Evgeniya Sukhodolskaya · February 20, 2026 In March 2025, we dropped a statement-bomb in the “Relevance Feedback in Information Retrieval” article and then went silent. We claimed that even though the information retrieval research field has proposed many useful mechanisms for increasing the relevance of search results, none of them made it to the neural search industry, simply because these approaches are not scalable. Certainly, there are methods widely used to improve the relevance of retrieved results: query rewriting, for example. Yet none of the vector search solutions out there have tried to use the possibilities that come with full access to the vector search index: traversing it in the direction of relevance, instead of guessing where to shoot the query to the vector space. To break this vector-search-native tools silence, we’re introducing Relevance Feedback Query, a universal, cheap and scalable method for improving the relevance recall of your search results. Vector Search Optimization In vector search, like in any field, you’re always trading off speed, cost, and quality on the way to production. Everyone needs instruments to tune that balance for their use case, like vector quantization to lower costs or reranking to increase relevance. For those instruments to work (and to stick in industry), they have to be universal, native to the stack, and reasonable. “Reasonable” means optimizing one side of the speed–cost–quality triangle without tanking the other two. At scale, neural search is usually limited to fairly simple dense encoders (embeddings up to a few thousand dimensions), as running large models to sift through billions of data points is, well, unreasonable. So most search quality-boosting tools aim to adjust/align/guide these simple retrievers. The guidance can come from the search end-user or a domain expert: rescoring via business-logic-based rules, reranking with a learned-to-rank model, and using relevance feedback mechanisms. What is Relevance Feedback Relevance feedback distills signals from current search results into the next retrieval iteration to surface more relevant documents. It can be provided by a human or a model, be binary (relevant–irrelevant) or granular, rescoring top documents by their relative relevance to the query. The same ranked results labeled by three kinds of feedback: pseudo-relevance, binary feedback, and re-scored feedback. Scores are illustrative. Based on this feedback, one of the retrieval components is adjusted: the query or the scoring method between the query and documents. The next retrieval iteration is done with this aligned component. For a detailed taxonomy with various methods from the field, check out our article “Relevance Feedback in Information Retrieval”. Relevance-feedback-based methods are a standard in full-text search, some of them proposed more than 50 years ago. Yet it feels like, when it comes to modern vector search, users have to reinvent the relevance feedback wheel due to the lack of universal interfaces: prompt a search agent to rewrite queries in endless, costly loops, do heuristics-based vector math and fine-tune models per request on the client side… Simply put, they have to struggle. Tools of a Vector Search Engine The goal is to fill this gap and let search end-users (doesn’t matter if humans or agents) get a proper relevance feedback interface to use in our vector search engine. So what makes a method a production-ready for vector search? …Should Be Cheap Since additional retrieval iterations already add to the search latency, we can’t afford to spend too much time or money on gathering feedback. That means: Very few documents for feedback The set of documents we pass to a feedback model should be extremely small. It is rarely affordable to run an LLM, use a cross encoder, or fine tune a machine learning model on hundreds of documents per request. A feedback model here is anything – agent, ML model, formula, … – that’s capable of providing (numerical) information on how relevant the retrieved document is to the query. We should already improve recall of the next retrieval iteration with that small sample. We can’t afford having the whole dataset as a feedback context for an agent. Automated You might have noticed that we wrote “a feedback model”. Humans are known to be the laziest at providing feedback, so the search feedback loop must run on its own. No labeling required Models which require many labels or domain expert input are hard to adopt. The feedback tool should be easy to train, and data – self-supervised. …Should Be Universal For every type of data Vector search is attractive because it is agnostic to the data type: images, video, audio, molecules… Not only text. Vector search native tools should be the same, operating on vectors without minding their nature. That is why query rewriting as reformulating text won’t suffice. For every type of signal Some relevance feedback methods depend on clean feedback signals: strictly relevant or irrelevant documents. In real life, especially when we need relevance boosting, results returned to a request could be, for example, meh and more meh. Ideally, the method should be able to use the noisy signals. Our Idea Usually neither LLMs nor subject matter experts providing feedback know the dataset’s exact shape or the vector space it spans. For that they would need to process a large subset of stored vectors. What they can do is help define a relevance direction from the examples they have seen. A retriever could then use these direction hints on the next retrieval iteration. Relevance Direction Imagine a hiker lost in a forest with a weak phone signal. They still manage to send a couple of quick photos to a friend who knows orienteering. The friend does not have a map of the whole forest or a live view of every tree. Still, they try to help and reply: “Better go away from anything looking like this plant. It might signal that a swamp is near.” “I see moss on this tree. Moss grows on the north side. If you keep north you might find a way out.” Now the hiker has extra knowledge to choose a path, comparing the landscape around them to the photos and applying the direction hints from their experienced friend. Read this as vector search. The forest is a vector space. The hiker is a retriever model. The brilliant friend is a feedback model. The photos are the hiker’s context, limited to the amount that is affordable to use for feedback. And the hiker’s new “where-should-I-go” decisions based on the acquired information are feedback-based (forest path) scoring. Using Entire Vector Space So the idea is to use feedback to define a more relevant direction (closer to this, further from that) within the vector space spanned by the dataset. That implies warping the notion of “closer” and “further”, the distance (or similarity) scoring metric used during the retrieval. Tweaking the relevance scoring formula based on the feedback model’s signal is nothing revolutionary. What changes is that this feedback-based scoring is used during the traversal of the entire dataset used by the vector search engine, not just a subset of results. Sketch of the Method So, stepping away from analogies (into an even deeper forest), the algorithm should consist of: Initial retrieval. Getting a small amount of feedback (within the context limit) on the results of this initial retrieval. A context limit is how many top documents from the initial retrieval the feedback model scores. We want to keep it as small as possible to save time and resources. A good choice for feedback format is granular judgments, aka pointwise relevance scores, between the query and the retrieved documents. They are more helpful when the retrieved documents are not strictly relevant or irrelevant. Extracting relevance signals from this feedback and propagating them into a new similarity (distance) scoring formula. Using this new feedback-based similarity scoring formula in the next retrieval iteration on the whole dataset of documents to traverse the vector index in the direction of more relevance. Simply put, we’re no longer using cosine to score similarities, but rather a formula adjusted for feedback. Dissecting Feedback Let’s talk about how a very small amount of feedback can be used most effectively, what relevance signals we can extract from it, and how exactly. Context Pairs If we show only one document to the feedback model, it is hard for it to say whether it is already a good match. Perhaps there are no relevant matches in the dataset at all. With two documents, this model could already judge which one is closer to what the user expects. Feedback then could be a context pair of (more) positive and (more) negative examples-documents. A context pair (positive, negative) is two documents from the top context limit results of the initial retrieval. The positive received a higher relevance-to-query score from the feedback model, and the negative received a lower score. Step through the example below: it picks a context pair, measures its confidence, computes a candidate’s delta, and combines them into a score. From feedback scores to a new score: the context pair, its confidence, a candidate’s delta, and the feedback-based score. All numbers are illustrative. Context Pair’s Confidence Two documents nearly indistinguishable from the perspective of a feedback model give far less information than a context pair with one clearly more relevant document. When documents in the context pair seem different to the feedback model, it is more confident in guiding the retriever. Direction’s Delta Then, if the retriever favors a document closer to the negative example by some delta, the retriever should most probably adjust its search direction in the vector space. Feedback-Based Scoring With the context pair(s) at our expense, we can try the following feedback-based scoring during retrieval: Let the retriever still have a say in what is relevant to the query, count in its score (to query). Yet reward candidates that are closer to the positive element of the context pair based on delta. Especially when the feedback model had high confidence in this pair. So, we need to combine signals (score, delta, confidence) in a reasonable, simple scoring formula with a few parameters. Usually what helps in coming up with a reasonable formula is looking at edge cases. The Math of Edge Cases Direction’s delta is zero If the retriever sees the positive and the negative documents (from the context pair) as equally similar to the candidate document, there’s no direction-establishing signal. Context pair’s confidence is zero. If, from the feedback model’s perspective, both documents in a context pair are identical to the query, there’s, once again, no information on direction of relevance. In both cases the final formula should rely on the retriever’s judgment, as opposed to a situation where feedback signals are strong. Naive Formula One of the options to express this behavior is through a weighted sum of signals, ensuring confidence and delta don’t collapse into a single joint term by exponentiating one of them. Applying the math of edge cases, we came up with this three-parameter (a, b and c) formula. $$ F = a \cdot \text{score} + \sum_{p=1}^{\text{# pairs}} \text{confidence}_{p}^{b} \cdot c \cdot \text{delta}_p $$ It computes the score between the query and a candidate document on the retrieval-with-relevance-feedback iteration. As you see, the amount of context pairs used in the scoring formula can be more than one but one is also an option. Is this formula set in s [truncated for AI cost control]