[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-qdrant-qdrant-relevance-feedback":3,"mdc--mxc9jt-key":48,"related-org-qdrant-qdrant-relevance-feedback":761,"related-repo-qdrant-qdrant-relevance-feedback":915},{"slug":4,"name":4,"fn":5,"description":6,"org":7,"tags":11,"stars":23,"repoUrl":24,"updatedAt":25,"license":26,"forks":27,"topics":28,"repo":43,"sourceUrl":46,"mdContent":47},"qdrant-relevance-feedback","optimize search results with relevance feedback","Expanding the candidate pool via relevance feedback, as an alternative to reranking when a dense retriever is too weak. Use when someone asks about 'Qdrant's Relevance Feedback API', 'improving dense search relevance\u002Frecall', 'how to discover\u002Fget more relevant results from vector search', 'cheaper\u002Fbetter alternative to reranking', 'using a more heavy\u002Fbig embedding model for dense search but can't afford it', 'finding more relevant documents beyond the initial search pool', or 'feedback loops'. Also trigger when the user has a search quality problem due to a dense retriever being weak and is considering reranking as a solution — this API may be a better fit",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},"qdrant","Qdrant","https:\u002F\u002Fpexgzepcugksgbtrxkhf.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Forg-logos\u002Fqdrant.png",[12,16,17,20],{"name":13,"slug":14,"type":15},"Performance","performance","tag",{"name":9,"slug":8,"type":15},{"name":18,"slug":19,"type":15},"Database","database",{"name":21,"slug":22,"type":15},"Search","search",197,"https:\u002F\u002Fgithub.com\u002Fqdrant\u002Fskills","2026-07-16T06:03:00.067171",null,23,[29,30,31,32,33,34,35,36,37,14,8,38,39,40,41,42],"agent-skills","ai-agents","claude-code","codex","cursor","embeddings","hybrid-search","monitoring","multitenancy","quantization","scaling","search-quality","vector-database","vector-search",{"repoUrl":24,"stars":23,"forks":27,"topics":44,"description":45},[29,30,31,32,33,34,35,36,37,14,8,38,39,40,41,42],"Agent skills for Qdrant vector search: scaling, performance optimization, search quality, monitoring, deployment, model migration, version upgrades, and SDK usage across Python, TypeScript, Rust, Go, .NET, Java","https:\u002F\u002Fgithub.com\u002Fqdrant\u002Fskills\u002Ftree\u002FHEAD\u002Fskills\u002Fqdrant-search-quality\u002Fsearch-strategies\u002Frelevance-feedback","---\nname: qdrant-relevance-feedback\ndescription: \"Expanding the candidate pool via relevance feedback, as an alternative to reranking when a dense retriever is too weak. Use when someone asks about 'Qdrant's Relevance Feedback API', 'improving dense search relevance\u002Frecall', 'how to discover\u002Fget more relevant results from vector search', 'cheaper\u002Fbetter alternative to reranking', 'using a more heavy\u002Fbig embedding model for dense search but can't afford it', 'finding more relevant documents beyond the initial search pool', or 'feedback loops'. Also trigger when the user has a search quality problem due to a dense retriever being weak and is considering reranking as a solution — this API may be a better fit\"\n---\n\nReranking reorders documents that have already been retrieved. Qdrant's Relevance Feedback (RF) instead modifies the vector search process itself based on a small amount of reranker feedback, distilling reranker (feedback model) knowledge into the search step. This allows RF to surface documents that the initial ANN search did not score highly enough.\n\nThe RF is intended for tasks where relevance correlates with similarity in vector space.\n\nHow you apply the RF depends on your goals.  \nFirst, understand how the RF works, read the ENTIRE section. Then define your goals and choose the appropriate usage pattern described below. Make sure to avoid the listed anti-patterns (\"DO NOTs\"). Before implementing anything, read CAREFULLY to avoid missing important details.\n\n## How It Works\n\nThe [Qdrant Query Point API with a type RelevanceFeedbackQuery](https:\u002F\u002Fskills.qdrant.tech\u002Fapi-reference\u002Fsearch\u002Fquery-points.md) takes:\n\n- a query (`target`)\n- a small list of seed documents (`feedback`) with relevance scores (often 4–5 seeds are enough)\n- formula weights, which MUST be trained once per general search use case (your dataset, dense retriever, and feedback model)\n\nIf you do not train the formula weights, results will at best be random, will not align with your data distribution or model behavior. Training is lightweight because the formula itself is simple.\n\nDuring search, it scores each candidate by combining similarity to the original query, similarity to highly rated seed documents and dissimilarity to poorly rated ones.\n\n### Feedback Model\n\nA **feedback model** is any model that can produce a float relevance score for `(query, document)` pairs. Higher scores must always mean higher relevance.\n\nExamples: a cross-encoder, embedding similarity (for example, cosine similarity between query and document embeddings, or max_sim for late interaction models), an LLM-based scorer, a custom ranker.\n\nThe feedback model used during training and inference MUST be the same model. Formula weights during training are calibrated to that model's score distribution. If you switch feedback models, you must retrain.\n\n**What is a Good Feedback Model:**\n- If the model does not improve ranking quality when used as a reranker on retrieved documents,  the RF search will not have a meaningful signal to amplify.\n- RF search quality depends heavily on how well the feedback model scores partial matches. The training loss of RF formula relies on relative ordering, so poor score separation in the middle range (documents that are neither clearly relevant nor clearly irrelevant) weakens results.\n\n### To Make the RF API Work, You Need to Calibrate Weights First\n\nUse when: setting up RF for a new use case — a new collection, feedback model, or embedding model powering ANN search.\n\nRF uses a weighted formula that combines the original query vector with feedback signals.\n\nFor the currently available `naive` strategy, the learned weights control:\n- `a` — how much to trust the original ANN query-document similarity\n- `b` — how strongly differences in feedback scores matter\n- `c` — how strongly to follow the feedback direction (toward relevant documents and away from irrelevant ones)\n\nThese weights must be learned from your data before use. You cannot safely use arbitrary values.\n\n- Install the [qdrant-relevance-feedback](https:\u002F\u002Fpypi.org\u002Fproject\u002Fqdrant-relevance-feedback\u002F) Python library. Study what goes into RelevanceFeedback.\n- Initialize a `RelevanceFeedback` instance. You can use provided QdrantRetriever or FastembedFeedback, or define your own.\n- Review `train` parameters before calling `train`. The library retrieves `limit` candidates per train query, scores them with the feedback model, learns the weighting parameters, and returns the calibrated values.\n- Call `train` on 50–200 representative, real, non-synthetic queries.\n  - Generate train queries yourself based on the use case, but give the option to the user to provide them, too.\n  - Inform user on cost and quality trade-offs of training.\n- Check train metrics which show if RF had a signal  (disagreement between retriever and feedback model) to distill and learn from. If there was no signal to learn from, adapt training parameters, queries or change a feedback model and retrain until RF learns well. \n- Store the resulting RF parameters in your configuration and use them during inference. Retrain if your query distribution or corpus changes significantly.\n- Evaluate resulting formula with `Evaluator` on a separate test set of representative, real, non-synthetic queries. If results seem unsatisfactory, investigate and inform user.  \n\nThe retriever, feedback model, and related parameters defined during training are assumed to remain the same during inference.\n\n## Want High-Quality Top-1\u002F3 Results at Reasonable Cost\n\nUse when: top-1 or top-3 precision matters most, and reranking a large pool of documents would be too expensive or slow. This pattern below can match reranking quality at the top of the ranking for semantic similarity tasks, but it performs worse at deeper cutoffs. Do not use this approach when top-10+ recall is the priority.\n\nOnly score a small set of seed documents. Five seeds is a robust default across many task types and scoring them costs user roughly 5× less than reranking a 25-document pool.\n\n- Retrieve the top 5 documents using ANN search. These become the feedback seeds. You'll need their stored embeddings.\n- Score them with the feedback model used in training.\n- Call Qdrant's Query API using the relevance feedback query:\n  - set `target` to the query retriever embedding (also possible to use Qdrant Cloud Inference).\n  - set `feedback` to a list of items where each item contains:\n    - `example=\u003Cseed vector, same embedding model as for `target`>` (also possible to use Qdrant Cloud Inference)\n    - `score=\u003Cfeedback model score>`\n  - set `using` to retriever's handle, RF operates in retriever's vector space. \n  - set `strategy` to `naive` with your calibrated parameters\n  - set `limit` to the number of final results you need and use the RF results directly as final results.\n\n Check the [Relevance Feedback Query API  documentation](https:\u002F\u002Fskills.qdrant.tech\u002Fmd\u002Fdocumentation\u002Fsearch\u002Fsearch-relevance\u002F?s=relevance-feedback) and study code\u002Fmethods of the relevant SDK before filling in anything.\n\nUsing a point ID in `example` causes the RF API to automatically exclude that document from the final results. Using stored embeddings used for retrieval instead potentially keeps the document in the final results.\n\n## Want to Find Relevant Documents Beyond the Initial Search Results\n\nUse when: recall matters more than latency or cost (research, legal, medical, compliance), and relevant documents may exist outside the initial ANN retrieval pool.\n\nIt performs two feedback model scoring rounds:\n1. on feedback seeds\n2. on newly surfaced RF results\n\nThe second reranking pass safely promotes newly discovered documents into the top-10 of the final ranking. The advantage over standard reranking is that RF can reach relevant documents that lie completely outside the initial ANN pool, while a reranker with the same budget cannot. The tradeoff is higher latency due to two rounds of feedback-model scoring.\n\n- Retrieve and score 5 seed documents. These become the feedback seeds. You'll need their point IDs.\n- Call Qdrant's query API using the relevance feedback query:\n  - set `target` to the query retriever embedding (also possible to use Qdrant Cloud Inference)\n  - set `feedback` to a list of items where each item contains:\n    - `example=\u003Cseed point ID>`\n    - `score=\u003Cfeedback score>`\n  - set `using` to retriever's handle, RF operates in retriever's vector space. \n  - set `strategy` to `naive` with your calibrated parameters\n  - set `limit` to the number of results user can afford to rerank based on the available cost budget. The total scoring cost equals the cost of scoring both the seeds and the RF results, roughly equivalent to reranking a pool of the same combined size. Inform and consult with the user.\n  - score the returned RF results with your feedback model.\n- Merge the original seeds and RF results, then sort by feedback score. These will be your final results.\n\n Check the [Relevance Feedback Query API  documentation](https:\u002F\u002Fskills.qdrant.tech\u002Fmd\u002Fdocumentation\u002Fsearch\u002Fsearch-relevance\u002F?s=relevance-feedback) and study code\u002Fmethods of the relevant SDK before filling in anything.\n\nUsing a point ID in `example` causes the RF API to automatically exclude that document from the final results. Using stored embeddings used for retrieval instead potentially keeps the document in the final results.\n\n## What NOT to Do\n\n- Do not skip calibration and use random formula weights. Untrained weights produce arbitrary results. (`a=1, b=0, c=0` can be used if you only want vanilla ANN behavior through the RF API.)\n- Do not use the RF API on sparse vectors.\n- Do not use a feedback model where higher scores mean lower relevance. Scores must be monotonic: higher = more relevant.\n- Do not use fewer than 2 feedback seeds. A single seed provides no contrastive signal. The formula needs at least one relatively more relevant and one relatively less relevant example to establish direction. Two is the minimum; five is the recommended default.\n- Do not use significantly more than 5 seeds expecting better quality. Additional seeds usually add noise and increase scoring cost without meaningful gains.\n- Do not use a different feedback model during inference than the one used during calibration. The learned weights are tied to that model's score scale and distribution.\n- Do not use a feedback model that does not improve retrieval quality as a standard reranker on your data.\n- Do not proceed to inference if training and evaluation metrics of qdrant-relevance-feedback package demonstrated unsatisfactory results, instead find a good training set of representative queries, a feedback model providing a meaningful signal and effective train parameters.\n",{"data":49,"body":50},{"name":4,"description":6},{"type":51,"children":52},"root",[53,61,66,77,84,100,137,142,147,154,175,180,185,193,206,212,217,222,235,270,275,380,385,391,396,401,517,531,544,550,555,560,574,579,684,694,704,710],{"type":54,"tag":55,"props":56,"children":57},"element","p",{},[58],{"type":59,"value":60},"text","Reranking reorders documents that have already been retrieved. Qdrant's Relevance Feedback (RF) instead modifies the vector search process itself based on a small amount of reranker feedback, distilling reranker (feedback model) knowledge into the search step. This allows RF to surface documents that the initial ANN search did not score highly enough.",{"type":54,"tag":55,"props":62,"children":63},{},[64],{"type":59,"value":65},"The RF is intended for tasks where relevance correlates with similarity in vector space.",{"type":54,"tag":55,"props":67,"children":68},{},[69,71,75],{"type":59,"value":70},"How you apply the RF depends on your goals.",{"type":54,"tag":72,"props":73,"children":74},"br",{},[],{"type":59,"value":76},"\nFirst, understand how the RF works, read the ENTIRE section. Then define your goals and choose the appropriate usage pattern described below. Make sure to avoid the listed anti-patterns (\"DO NOTs\"). Before implementing anything, read CAREFULLY to avoid missing important details.",{"type":54,"tag":78,"props":79,"children":81},"h2",{"id":80},"how-it-works",[82],{"type":59,"value":83},"How It Works",{"type":54,"tag":55,"props":85,"children":86},{},[87,89,98],{"type":59,"value":88},"The ",{"type":54,"tag":90,"props":91,"children":95},"a",{"href":92,"rel":93},"https:\u002F\u002Fskills.qdrant.tech\u002Fapi-reference\u002Fsearch\u002Fquery-points.md",[94],"nofollow",[96],{"type":59,"value":97},"Qdrant Query Point API with a type RelevanceFeedbackQuery",{"type":59,"value":99}," takes:",{"type":54,"tag":101,"props":102,"children":103},"ul",{},[104,119,132],{"type":54,"tag":105,"props":106,"children":107},"li",{},[108,110,117],{"type":59,"value":109},"a query (",{"type":54,"tag":111,"props":112,"children":114},"code",{"className":113},[],[115],{"type":59,"value":116},"target",{"type":59,"value":118},")",{"type":54,"tag":105,"props":120,"children":121},{},[122,124,130],{"type":59,"value":123},"a small list of seed documents (",{"type":54,"tag":111,"props":125,"children":127},{"className":126},[],[128],{"type":59,"value":129},"feedback",{"type":59,"value":131},") with relevance scores (often 4–5 seeds are enough)",{"type":54,"tag":105,"props":133,"children":134},{},[135],{"type":59,"value":136},"formula weights, which MUST be trained once per general search use case (your dataset, dense retriever, and feedback model)",{"type":54,"tag":55,"props":138,"children":139},{},[140],{"type":59,"value":141},"If you do not train the formula weights, results will at best be random, will not align with your data distribution or model behavior. Training is lightweight because the formula itself is simple.",{"type":54,"tag":55,"props":143,"children":144},{},[145],{"type":59,"value":146},"During search, it scores each candidate by combining similarity to the original query, similarity to highly rated seed documents and dissimilarity to poorly rated ones.",{"type":54,"tag":148,"props":149,"children":151},"h3",{"id":150},"feedback-model",[152],{"type":59,"value":153},"Feedback Model",{"type":54,"tag":55,"props":155,"children":156},{},[157,159,165,167,173],{"type":59,"value":158},"A ",{"type":54,"tag":160,"props":161,"children":162},"strong",{},[163],{"type":59,"value":164},"feedback model",{"type":59,"value":166}," is any model that can produce a float relevance score for ",{"type":54,"tag":111,"props":168,"children":170},{"className":169},[],[171],{"type":59,"value":172},"(query, document)",{"type":59,"value":174}," pairs. Higher scores must always mean higher relevance.",{"type":54,"tag":55,"props":176,"children":177},{},[178],{"type":59,"value":179},"Examples: a cross-encoder, embedding similarity (for example, cosine similarity between query and document embeddings, or max_sim for late interaction models), an LLM-based scorer, a custom ranker.",{"type":54,"tag":55,"props":181,"children":182},{},[183],{"type":59,"value":184},"The feedback model used during training and inference MUST be the same model. Formula weights during training are calibrated to that model's score distribution. If you switch feedback models, you must retrain.",{"type":54,"tag":55,"props":186,"children":187},{},[188],{"type":54,"tag":160,"props":189,"children":190},{},[191],{"type":59,"value":192},"What is a Good Feedback Model:",{"type":54,"tag":101,"props":194,"children":195},{},[196,201],{"type":54,"tag":105,"props":197,"children":198},{},[199],{"type":59,"value":200},"If the model does not improve ranking quality when used as a reranker on retrieved documents,  the RF search will not have a meaningful signal to amplify.",{"type":54,"tag":105,"props":202,"children":203},{},[204],{"type":59,"value":205},"RF search quality depends heavily on how well the feedback model scores partial matches. The training loss of RF formula relies on relative ordering, so poor score separation in the middle range (documents that are neither clearly relevant nor clearly irrelevant) weakens results.",{"type":54,"tag":148,"props":207,"children":209},{"id":208},"to-make-the-rf-api-work-you-need-to-calibrate-weights-first",[210],{"type":59,"value":211},"To Make the RF API Work, You Need to Calibrate Weights First",{"type":54,"tag":55,"props":213,"children":214},{},[215],{"type":59,"value":216},"Use when: setting up RF for a new use case — a new collection, feedback model, or embedding model powering ANN search.",{"type":54,"tag":55,"props":218,"children":219},{},[220],{"type":59,"value":221},"RF uses a weighted formula that combines the original query vector with feedback signals.",{"type":54,"tag":55,"props":223,"children":224},{},[225,227,233],{"type":59,"value":226},"For the currently available ",{"type":54,"tag":111,"props":228,"children":230},{"className":229},[],[231],{"type":59,"value":232},"naive",{"type":59,"value":234}," strategy, the learned weights control:",{"type":54,"tag":101,"props":236,"children":237},{},[238,248,259],{"type":54,"tag":105,"props":239,"children":240},{},[241,246],{"type":54,"tag":111,"props":242,"children":244},{"className":243},[],[245],{"type":59,"value":90},{"type":59,"value":247}," — how much to trust the original ANN query-document similarity",{"type":54,"tag":105,"props":249,"children":250},{},[251,257],{"type":54,"tag":111,"props":252,"children":254},{"className":253},[],[255],{"type":59,"value":256},"b",{"type":59,"value":258}," — how strongly differences in feedback scores matter",{"type":54,"tag":105,"props":260,"children":261},{},[262,268],{"type":54,"tag":111,"props":263,"children":265},{"className":264},[],[266],{"type":59,"value":267},"c",{"type":59,"value":269}," — how strongly to follow the feedback direction (toward relevant documents and away from irrelevant ones)",{"type":54,"tag":55,"props":271,"children":272},{},[273],{"type":59,"value":274},"These weights must be learned from your data before use. You cannot safely use arbitrary values.",{"type":54,"tag":101,"props":276,"children":277},{},[278,291,304,332,357,362,367],{"type":54,"tag":105,"props":279,"children":280},{},[281,283,289],{"type":59,"value":282},"Install the ",{"type":54,"tag":90,"props":284,"children":287},{"href":285,"rel":286},"https:\u002F\u002Fpypi.org\u002Fproject\u002Fqdrant-relevance-feedback\u002F",[94],[288],{"type":59,"value":4},{"type":59,"value":290}," Python library. Study what goes into RelevanceFeedback.",{"type":54,"tag":105,"props":292,"children":293},{},[294,296,302],{"type":59,"value":295},"Initialize a ",{"type":54,"tag":111,"props":297,"children":299},{"className":298},[],[300],{"type":59,"value":301},"RelevanceFeedback",{"type":59,"value":303}," instance. You can use provided QdrantRetriever or FastembedFeedback, or define your own.",{"type":54,"tag":105,"props":305,"children":306},{},[307,309,315,317,322,324,330],{"type":59,"value":308},"Review ",{"type":54,"tag":111,"props":310,"children":312},{"className":311},[],[313],{"type":59,"value":314},"train",{"type":59,"value":316}," parameters before calling ",{"type":54,"tag":111,"props":318,"children":320},{"className":319},[],[321],{"type":59,"value":314},{"type":59,"value":323},". The library retrieves ",{"type":54,"tag":111,"props":325,"children":327},{"className":326},[],[328],{"type":59,"value":329},"limit",{"type":59,"value":331}," candidates per train query, scores them with the feedback model, learns the weighting parameters, and returns the calibrated values.",{"type":54,"tag":105,"props":333,"children":334},{},[335,337,342,344],{"type":59,"value":336},"Call ",{"type":54,"tag":111,"props":338,"children":340},{"className":339},[],[341],{"type":59,"value":314},{"type":59,"value":343}," on 50–200 representative, real, non-synthetic queries.\n",{"type":54,"tag":101,"props":345,"children":346},{},[347,352],{"type":54,"tag":105,"props":348,"children":349},{},[350],{"type":59,"value":351},"Generate train queries yourself based on the use case, but give the option to the user to provide them, too.",{"type":54,"tag":105,"props":353,"children":354},{},[355],{"type":59,"value":356},"Inform user on cost and quality trade-offs of training.",{"type":54,"tag":105,"props":358,"children":359},{},[360],{"type":59,"value":361},"Check train metrics which show if RF had a signal  (disagreement between retriever and feedback model) to distill and learn from. If there was no signal to learn from, adapt training parameters, queries or change a feedback model and retrain until RF learns well.",{"type":54,"tag":105,"props":363,"children":364},{},[365],{"type":59,"value":366},"Store the resulting RF parameters in your configuration and use them during inference. Retrain if your query distribution or corpus changes significantly.",{"type":54,"tag":105,"props":368,"children":369},{},[370,372,378],{"type":59,"value":371},"Evaluate resulting formula with ",{"type":54,"tag":111,"props":373,"children":375},{"className":374},[],[376],{"type":59,"value":377},"Evaluator",{"type":59,"value":379}," on a separate test set of representative, real, non-synthetic queries. If results seem unsatisfactory, investigate and inform user.",{"type":54,"tag":55,"props":381,"children":382},{},[383],{"type":59,"value":384},"The retriever, feedback model, and related parameters defined during training are assumed to remain the same during inference.",{"type":54,"tag":78,"props":386,"children":388},{"id":387},"want-high-quality-top-13-results-at-reasonable-cost",[389],{"type":59,"value":390},"Want High-Quality Top-1\u002F3 Results at Reasonable Cost",{"type":54,"tag":55,"props":392,"children":393},{},[394],{"type":59,"value":395},"Use when: top-1 or top-3 precision matters most, and reranking a large pool of documents would be too expensive or slow. This pattern below can match reranking quality at the top of the ranking for semantic similarity tasks, but it performs worse at deeper cutoffs. Do not use this approach when top-10+ recall is the priority.",{"type":54,"tag":55,"props":397,"children":398},{},[399],{"type":59,"value":400},"Only score a small set of seed documents. Five seeds is a robust default across many task types and scoring them costs user roughly 5× less than reranking a 25-document pool.",{"type":54,"tag":101,"props":402,"children":403},{},[404,409,414],{"type":54,"tag":105,"props":405,"children":406},{},[407],{"type":59,"value":408},"Retrieve the top 5 documents using ANN search. These become the feedback seeds. You'll need their stored embeddings.",{"type":54,"tag":105,"props":410,"children":411},{},[412],{"type":59,"value":413},"Score them with the feedback model used in training.",{"type":54,"tag":105,"props":415,"children":416},{},[417,419],{"type":59,"value":418},"Call Qdrant's Query API using the relevance feedback query:\n",{"type":54,"tag":101,"props":420,"children":421},{},[422,434,475,487,506],{"type":54,"tag":105,"props":423,"children":424},{},[425,427,432],{"type":59,"value":426},"set ",{"type":54,"tag":111,"props":428,"children":430},{"className":429},[],[431],{"type":59,"value":116},{"type":59,"value":433}," to the query retriever embedding (also possible to use Qdrant Cloud Inference).",{"type":54,"tag":105,"props":435,"children":436},{},[437,438,443,445],{"type":59,"value":426},{"type":54,"tag":111,"props":439,"children":441},{"className":440},[],[442],{"type":59,"value":129},{"type":59,"value":444}," to a list of items where each item contains:\n",{"type":54,"tag":101,"props":446,"children":447},{},[448,466],{"type":54,"tag":105,"props":449,"children":450},{},[451,457,458,464],{"type":54,"tag":111,"props":452,"children":454},{"className":453},[],[455],{"type":59,"value":456},"example=\u003Cseed vector, same embedding model as for ",{"type":59,"value":116},{"type":54,"tag":111,"props":459,"children":461},{"className":460},[],[462],{"type":59,"value":463},">",{"type":59,"value":465}," (also possible to use Qdrant Cloud Inference)",{"type":54,"tag":105,"props":467,"children":468},{},[469],{"type":54,"tag":111,"props":470,"children":472},{"className":471},[],[473],{"type":59,"value":474},"score=\u003Cfeedback model score>",{"type":54,"tag":105,"props":476,"children":477},{},[478,479,485],{"type":59,"value":426},{"type":54,"tag":111,"props":480,"children":482},{"className":481},[],[483],{"type":59,"value":484},"using",{"type":59,"value":486}," to retriever's handle, RF operates in retriever's vector space.",{"type":54,"tag":105,"props":488,"children":489},{},[490,491,497,499,504],{"type":59,"value":426},{"type":54,"tag":111,"props":492,"children":494},{"className":493},[],[495],{"type":59,"value":496},"strategy",{"type":59,"value":498}," to ",{"type":54,"tag":111,"props":500,"children":502},{"className":501},[],[503],{"type":59,"value":232},{"type":59,"value":505}," with your calibrated parameters",{"type":54,"tag":105,"props":507,"children":508},{},[509,510,515],{"type":59,"value":426},{"type":54,"tag":111,"props":511,"children":513},{"className":512},[],[514],{"type":59,"value":329},{"type":59,"value":516}," to the number of final results you need and use the RF results directly as final results.",{"type":54,"tag":55,"props":518,"children":519},{},[520,522,529],{"type":59,"value":521},"Check the ",{"type":54,"tag":90,"props":523,"children":526},{"href":524,"rel":525},"https:\u002F\u002Fskills.qdrant.tech\u002Fmd\u002Fdocumentation\u002Fsearch\u002Fsearch-relevance\u002F?s=relevance-feedback",[94],[527],{"type":59,"value":528},"Relevance Feedback Query API  documentation",{"type":59,"value":530}," and study code\u002Fmethods of the relevant SDK before filling in anything.",{"type":54,"tag":55,"props":532,"children":533},{},[534,536,542],{"type":59,"value":535},"Using a point ID in ",{"type":54,"tag":111,"props":537,"children":539},{"className":538},[],[540],{"type":59,"value":541},"example",{"type":59,"value":543}," causes the RF API to automatically exclude that document from the final results. Using stored embeddings used for retrieval instead potentially keeps the document in the final results.",{"type":54,"tag":78,"props":545,"children":547},{"id":546},"want-to-find-relevant-documents-beyond-the-initial-search-results",[548],{"type":59,"value":549},"Want to Find Relevant Documents Beyond the Initial Search Results",{"type":54,"tag":55,"props":551,"children":552},{},[553],{"type":59,"value":554},"Use when: recall matters more than latency or cost (research, legal, medical, compliance), and relevant documents may exist outside the initial ANN retrieval pool.",{"type":54,"tag":55,"props":556,"children":557},{},[558],{"type":59,"value":559},"It performs two feedback model scoring rounds:",{"type":54,"tag":561,"props":562,"children":563},"ol",{},[564,569],{"type":54,"tag":105,"props":565,"children":566},{},[567],{"type":59,"value":568},"on feedback seeds",{"type":54,"tag":105,"props":570,"children":571},{},[572],{"type":59,"value":573},"on newly surfaced RF results",{"type":54,"tag":55,"props":575,"children":576},{},[577],{"type":59,"value":578},"The second reranking pass safely promotes newly discovered documents into the top-10 of the final ranking. The advantage over standard reranking is that RF can reach relevant documents that lie completely outside the initial ANN pool, while a reranker with the same budget cannot. The tradeoff is higher latency due to two rounds of feedback-model scoring.",{"type":54,"tag":101,"props":580,"children":581},{},[582,587,679],{"type":54,"tag":105,"props":583,"children":584},{},[585],{"type":59,"value":586},"Retrieve and score 5 seed documents. These become the feedback seeds. You'll need their point IDs.",{"type":54,"tag":105,"props":588,"children":589},{},[590,592],{"type":59,"value":591},"Call Qdrant's query API using the relevance feedback query:\n",{"type":54,"tag":101,"props":593,"children":594},{},[595,606,637,647,663,674],{"type":54,"tag":105,"props":596,"children":597},{},[598,599,604],{"type":59,"value":426},{"type":54,"tag":111,"props":600,"children":602},{"className":601},[],[603],{"type":59,"value":116},{"type":59,"value":605}," to the query retriever embedding (also possible to use Qdrant Cloud Inference)",{"type":54,"tag":105,"props":607,"children":608},{},[609,610,615,616],{"type":59,"value":426},{"type":54,"tag":111,"props":611,"children":613},{"className":612},[],[614],{"type":59,"value":129},{"type":59,"value":444},{"type":54,"tag":101,"props":617,"children":618},{},[619,628],{"type":54,"tag":105,"props":620,"children":621},{},[622],{"type":54,"tag":111,"props":623,"children":625},{"className":624},[],[626],{"type":59,"value":627},"example=\u003Cseed point ID>",{"type":54,"tag":105,"props":629,"children":630},{},[631],{"type":54,"tag":111,"props":632,"children":634},{"className":633},[],[635],{"type":59,"value":636},"score=\u003Cfeedback score>",{"type":54,"tag":105,"props":638,"children":639},{},[640,641,646],{"type":59,"value":426},{"type":54,"tag":111,"props":642,"children":644},{"className":643},[],[645],{"type":59,"value":484},{"type":59,"value":486},{"type":54,"tag":105,"props":648,"children":649},{},[650,651,656,657,662],{"type":59,"value":426},{"type":54,"tag":111,"props":652,"children":654},{"className":653},[],[655],{"type":59,"value":496},{"type":59,"value":498},{"type":54,"tag":111,"props":658,"children":660},{"className":659},[],[661],{"type":59,"value":232},{"type":59,"value":505},{"type":54,"tag":105,"props":664,"children":665},{},[666,667,672],{"type":59,"value":426},{"type":54,"tag":111,"props":668,"children":670},{"className":669},[],[671],{"type":59,"value":329},{"type":59,"value":673}," to the number of results user can afford to rerank based on the available cost budget. The total scoring cost equals the cost of scoring both the seeds and the RF results, roughly equivalent to reranking a pool of the same combined size. Inform and consult with the user.",{"type":54,"tag":105,"props":675,"children":676},{},[677],{"type":59,"value":678},"score the returned RF results with your feedback model.",{"type":54,"tag":105,"props":680,"children":681},{},[682],{"type":59,"value":683},"Merge the original seeds and RF results, then sort by feedback score. These will be your final results.",{"type":54,"tag":55,"props":685,"children":686},{},[687,688,693],{"type":59,"value":521},{"type":54,"tag":90,"props":689,"children":691},{"href":524,"rel":690},[94],[692],{"type":59,"value":528},{"type":59,"value":530},{"type":54,"tag":55,"props":695,"children":696},{},[697,698,703],{"type":59,"value":535},{"type":54,"tag":111,"props":699,"children":701},{"className":700},[],[702],{"type":59,"value":541},{"type":59,"value":543},{"type":54,"tag":78,"props":705,"children":707},{"id":706},"what-not-to-do",[708],{"type":59,"value":709},"What NOT to Do",{"type":54,"tag":101,"props":711,"children":712},{},[713,726,731,736,741,746,751,756],{"type":54,"tag":105,"props":714,"children":715},{},[716,718,724],{"type":59,"value":717},"Do not skip calibration and use random formula weights. Untrained weights produce arbitrary results. (",{"type":54,"tag":111,"props":719,"children":721},{"className":720},[],[722],{"type":59,"value":723},"a=1, b=0, c=0",{"type":59,"value":725}," can be used if you only want vanilla ANN behavior through the RF API.)",{"type":54,"tag":105,"props":727,"children":728},{},[729],{"type":59,"value":730},"Do not use the RF API on sparse vectors.",{"type":54,"tag":105,"props":732,"children":733},{},[734],{"type":59,"value":735},"Do not use a feedback model where higher scores mean lower relevance. Scores must be monotonic: higher = more relevant.",{"type":54,"tag":105,"props":737,"children":738},{},[739],{"type":59,"value":740},"Do not use fewer than 2 feedback seeds. A single seed provides no contrastive signal. The formula needs at least one relatively more relevant and one relatively less relevant example to establish direction. Two is the minimum; five is the recommended default.",{"type":54,"tag":105,"props":742,"children":743},{},[744],{"type":59,"value":745},"Do not use significantly more than 5 seeds expecting better quality. Additional seeds usually add noise and increase scoring cost without meaningful gains.",{"type":54,"tag":105,"props":747,"children":748},{},[749],{"type":59,"value":750},"Do not use a different feedback model during inference than the one used during calibration. The learned weights are tied to that model's score scale and distribution.",{"type":54,"tag":105,"props":752,"children":753},{},[754],{"type":59,"value":755},"Do not use a feedback model that does not improve retrieval quality as a standard reranker on your data.",{"type":54,"tag":105,"props":757,"children":758},{},[759],{"type":59,"value":760},"Do not proceed to inference if training and evaluation metrics of qdrant-relevance-feedback package demonstrated unsatisfactory results, instead find a good training set of representative queries, a feedback model providing a meaningful signal and effective train parameters.",{"items":762,"total":914},[763,778,793,808,821,835,848,858,869,880,890,901],{"slug":764,"name":764,"fn":765,"description":766,"org":767,"tags":768,"stars":23,"repoUrl":24,"updatedAt":777},"qdrant-advisor","diagnose and troubleshoot Qdrant deployments","Diagnose, troubleshoot, and advise on any Qdrant deployment by loading the latest official Qdrant skills live from skills.qdrant.tech. Use this whenever someone raises a Qdrant problem or question — slow or degraded search, high or growing memory \u002F OOM crashes, optimizer stuck or slow, indexing slowness, scaling and sharding decisions (node count, QPS, latency, multitenancy, vertical vs horizontal), poor or irrelevant search results, hybrid search and reranking, embedding-model migration, version upgrades and compatibility, monitoring and observability (Prometheus, Grafana, health checks, \u002Fmetrics, \u002Ftelemetry), deployment choices (local, Docker, self-hosted, Qdrant Cloud, embedded), or client-SDK questions (Python, TypeScript, Rust, Go, .NET, Java). Trigger especially when the context is clearly a Qdrant cluster, collection, or vector-search deployment. Always prefer this skill over answering from memory: it pulls current, authoritative guidance and only the relevant context.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[769,770,773,776],{"name":18,"slug":19,"type":15},{"name":771,"slug":772,"type":15},"Debugging","debugging",{"name":774,"slug":775,"type":15},"Operations","operations",{"name":9,"slug":8,"type":15},"2026-07-16T06:02:56.675293",{"slug":779,"name":779,"fn":780,"description":781,"org":782,"tags":783,"stars":23,"repoUrl":24,"updatedAt":792},"qdrant-clients-sdk","integrate Qdrant client SDKs","Qdrant provides client SDKs for various programming languages, allowing easy integration with Qdrant deployments.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[784,787,788,789],{"name":785,"slug":786,"type":15},"API Development","api-development",{"name":18,"slug":19,"type":15},{"name":9,"slug":8,"type":15},{"name":790,"slug":791,"type":15},"SDK","sdk","2026-07-16T05:59:57.799789",{"slug":794,"name":794,"fn":795,"description":796,"org":797,"tags":798,"stars":23,"repoUrl":24,"updatedAt":807},"qdrant-deployment-options","select Qdrant deployment options","Guides Qdrant deployment selection. Use when someone asks 'how to deploy Qdrant', 'Docker vs Cloud', 'local mode', 'embedded Qdrant', 'Qdrant EDGE', 'which deployment option', 'self-hosted vs cloud', or 'need lowest latency deployment'. Also use when choosing between deployment types for a new project.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[799,800,803,806],{"name":18,"slug":19,"type":15},{"name":801,"slug":802,"type":15},"Deployment","deployment",{"name":804,"slug":805,"type":15},"Infrastructure","infrastructure",{"name":9,"slug":8,"type":15},"2026-07-19T05:38:33.324265",{"slug":809,"name":809,"fn":810,"description":811,"org":812,"tags":813,"stars":23,"repoUrl":24,"updatedAt":820},"qdrant-edge","build applications with Qdrant Edge","Guides building on Qdrant Edge, the embedded in-process shard. Use when someone asks 'how to sync Edge with the server', 'keep a local shard in sync with Qdrant Cloud', 'BM25 or keyword search on Edge', 'hybrid search on Edge', 'embeddings on device', 'Edge snapshots', 'apply a partial snapshot', 'why is my Edge search empty after inserts', or is writing custom sync, BM25, or fusion code against qdrant-edge. Also use when deciding what Edge ships built-in versus what you must implement.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[814,815,818,819],{"name":18,"slug":19,"type":15},{"name":816,"slug":817,"type":15},"Edge","edge",{"name":9,"slug":8,"type":15},{"name":21,"slug":22,"type":15},"2026-07-16T06:01:19.179105",{"slug":822,"name":822,"fn":823,"description":824,"org":825,"tags":826,"stars":23,"repoUrl":24,"updatedAt":834},"qdrant-horizontal-scaling","guide Qdrant horizontal scaling decisions","Diagnoses and guides Qdrant horizontal scaling decisions. Use when someone asks 'vertical or horizontal?', 'how many nodes?', 'how many shards?', 'how to add nodes', 'resharding', 'data doesn't fit', or 'need more capacity'. Also use when data growth outpaces current deployment.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[827,830,831,832],{"name":828,"slug":829,"type":15},"Architecture","architecture",{"name":18,"slug":19,"type":15},{"name":9,"slug":8,"type":15},{"name":833,"slug":39,"type":15},"Scaling","2026-07-19T05:38:32.348551",{"slug":836,"name":836,"fn":837,"description":838,"org":839,"tags":840,"stars":23,"repoUrl":24,"updatedAt":847},"qdrant-hybrid-search","implement hybrid search in Qdrant","Explains hybrid search in Qdrant. Use when someone asks 'how do I setup hybrid search?', 'how to combine keyword and semantic search?', 'sparse plus dense vectors?', 'missing keyword matches', 'how to combine results from multiple searches?' and 'combining multiple representations'",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[841,844,845,846],{"name":842,"slug":843,"type":15},"AI Infrastructure","ai-infrastructure",{"name":18,"slug":19,"type":15},{"name":9,"slug":8,"type":15},{"name":21,"slug":22,"type":15},"2026-07-16T06:00:12.328122",{"slug":849,"name":849,"fn":850,"description":851,"org":852,"tags":853,"stars":23,"repoUrl":24,"updatedAt":857},"qdrant-hybrid-search-combining","combine hybrid search scores in Qdrant","Fusing scores from multiple searches into a single ranked result (RRF, DBSF, custom fusion). Use when someone asks 'RRF or DBSF?', 'how to combine sparse and dense', 'how to combine scores from multiple searches?', 'custom fusion', or 'fusion is not producing good results'",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[854,855,856],{"name":18,"slug":19,"type":15},{"name":9,"slug":8,"type":15},{"name":21,"slug":22,"type":15},"2026-07-16T06:01:49.401142",{"slug":859,"name":859,"fn":860,"description":861,"org":862,"tags":863,"stars":23,"repoUrl":24,"updatedAt":868},"qdrant-hybrid-search-prefetches","configure hybrid search with prefetch queries","Constructing prefetch queries for hybrid retrieval, including sparse\u002Fdense and multi-field setups, and choosing a sparse embedding model. Use when someone asks 'dense and sparse in one search?', 'how to combine multiple fields for retrieval?', 'payloads or sparse vectors for lexical?', 'which sparse embedding model to use?', or 'BM25 vs SPLADE?'",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[864,865,866,867],{"name":785,"slug":786,"type":15},{"name":18,"slug":19,"type":15},{"name":9,"slug":8,"type":15},{"name":21,"slug":22,"type":15},"2026-07-16T06:03:02.305907",{"slug":870,"name":870,"fn":871,"description":872,"org":873,"tags":874,"stars":23,"repoUrl":24,"updatedAt":879},"qdrant-indexing-performance-optimization","optimize Qdrant indexing and data ingestion","Diagnoses and fixes slow Qdrant indexing and data ingestion. Use when someone reports 'uploads are slow', 'indexing takes forever', 'optimizer is stuck', 'HNSW build time too long', or 'data uploaded but search is bad'. Also use when optimizer status shows errors, segments won't merge, or indexing threshold questions arise.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[875,876,877,878],{"name":18,"slug":19,"type":15},{"name":771,"slug":772,"type":15},{"name":13,"slug":14,"type":15},{"name":9,"slug":8,"type":15},"2026-07-19T05:38:36.321999",{"slug":881,"name":881,"fn":882,"description":883,"org":884,"tags":885,"stars":23,"repoUrl":24,"updatedAt":889},"qdrant-memory-usage-optimization","optimize Qdrant memory usage","Diagnoses and reduces Qdrant memory usage. Use when someone reports 'memory too high', 'RAM keeps growing', 'node crashed', 'out of memory', 'memory leak', or asks 'why is memory usage so high?', 'how to reduce RAM?'. Also use when memory doesn't match calculations, quantization didn't help, or nodes crash during recovery.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[886,887,888],{"name":771,"slug":772,"type":15},{"name":13,"slug":14,"type":15},{"name":9,"slug":8,"type":15},"2026-07-16T06:00:15.357243",{"slug":891,"name":891,"fn":892,"description":893,"org":894,"tags":895,"stars":23,"repoUrl":24,"updatedAt":900},"qdrant-minimize-latency","optimize Qdrant query latency","Guides Qdrant query latency optimization. Use when someone asks 'search is slow', 'how to reduce latency', 'p99 is too high', 'tail latency', 'single query too slow', 'how to make search faster', or 'latency spikes'.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[896,897,898,899],{"name":18,"slug":19,"type":15},{"name":771,"slug":772,"type":15},{"name":13,"slug":14,"type":15},{"name":9,"slug":8,"type":15},"2026-07-16T05:59:17.325839",{"slug":902,"name":902,"fn":903,"description":904,"org":905,"tags":906,"stars":23,"repoUrl":24,"updatedAt":913},"qdrant-model-migration","migrate embedding models in Qdrant","Guides embedding model migration in Qdrant without downtime. Use when someone asks 'how to switch embedding models', 'how to migrate vectors', 'how to update to a new model', 'zero-downtime model change', 'how to re-embed my data', or 'can I use two models at once'. Also use when upgrading model dimensions, switching providers, or A\u002FB testing models.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[907,908,911,912],{"name":18,"slug":19,"type":15},{"name":909,"slug":910,"type":15},"Migration","migration",{"name":13,"slug":14,"type":15},{"name":9,"slug":8,"type":15},"2026-07-16T06:00:48.841055",30,{"items":916,"total":914},[917,924,931,938,945,952,959],{"slug":764,"name":764,"fn":765,"description":766,"org":918,"tags":919,"stars":23,"repoUrl":24,"updatedAt":777},{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[920,921,922,923],{"name":18,"slug":19,"type":15},{"name":771,"slug":772,"type":15},{"name":774,"slug":775,"type":15},{"name":9,"slug":8,"type":15},{"slug":779,"name":779,"fn":780,"description":781,"org":925,"tags":926,"stars":23,"repoUrl":24,"updatedAt":792},{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[927,928,929,930],{"name":785,"slug":786,"type":15},{"name":18,"slug":19,"type":15},{"name":9,"slug":8,"type":15},{"name":790,"slug":791,"type":15},{"slug":794,"name":794,"fn":795,"description":796,"org":932,"tags":933,"stars":23,"repoUrl":24,"updatedAt":807},{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[934,935,936,937],{"name":18,"slug":19,"type":15},{"name":801,"slug":802,"type":15},{"name":804,"slug":805,"type":15},{"name":9,"slug":8,"type":15},{"slug":809,"name":809,"fn":810,"description":811,"org":939,"tags":940,"stars":23,"repoUrl":24,"updatedAt":820},{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[941,942,943,944],{"name":18,"slug":19,"type":15},{"name":816,"slug":817,"type":15},{"name":9,"slug":8,"type":15},{"name":21,"slug":22,"type":15},{"slug":822,"name":822,"fn":823,"description":824,"org":946,"tags":947,"stars":23,"repoUrl":24,"updatedAt":834},{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[948,949,950,951],{"name":828,"slug":829,"type":15},{"name":18,"slug":19,"type":15},{"name":9,"slug":8,"type":15},{"name":833,"slug":39,"type":15},{"slug":836,"name":836,"fn":837,"description":838,"org":953,"tags":954,"stars":23,"repoUrl":24,"updatedAt":847},{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[955,956,957,958],{"name":842,"slug":843,"type":15},{"name":18,"slug":19,"type":15},{"name":9,"slug":8,"type":15},{"name":21,"slug":22,"type":15},{"slug":849,"name":849,"fn":850,"description":851,"org":960,"tags":961,"stars":23,"repoUrl":24,"updatedAt":857},{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[962,963,964],{"name":18,"slug":19,"type":15},{"name":9,"slug":8,"type":15},{"name":21,"slug":22,"type":15}]