Recipes
Production-ready workflows that pair the right extractors with the right retrievers for common use cases.
50 recipes available
Feature Extraction
Turn raw media into structured intelligence
Semantic Multimodal Search
Find anything across video, image, audio, and documents
Hierarchical Classification
Auto-label content into structured taxonomies
Multimodal RAG
LLMs that cite real clips, frames, and documents
Clustering & Theme Discovery
Reveal structure you didn't know existed
Semantic Join
Link extracted content to business context
Anomaly Detection
Spot unusual content automatically
Dataset Versioning
Rebuild any dataset state deterministically
Semantic Drift Detection
Track how your data changes over time
Sports Highlights Pipeline
Auto-detect and assemble highlight moments from sports footage
BYO Embeddings Vector Search
Bring your own embeddings and search in 60 seconds
Dense Search Over Your Own Embeddings, and What Hybrid Needs
Bring your own vectors to MVS, search them by vector, and know the one setup step BM25 needs
What are Mixpeek Recipes?
Mixpeek recipes are practical blueprints for multimodal retrieval pipelines. They demonstrate how to combine feature extractors, retriever stages, and enrichment resources to solve real ML problems.
Composable Patterns
Each recipe shows how to combine extractors, stages, and enrichment resources. Copy the pattern and customize for your use case.
ML-Native Workflows
Semantic search, anomaly detection, dataset engineering-recipes are organized by the ML patterns you're trying to implement.
Production Ready
Clone templates directly into Mixpeek Studio. Each recipe includes Python snippets, retriever stages, and enrichment configurations.
Recent updates
Full changelog- Sep 16, 2026A recipe for running a large library without paying twice for the same fileFour of the questions on the solution finder were about volume, cost, continuous ingest and search latency, and all four pointed at a recipe about what to extract rather than how to do it at scale. There is now one that answers them. It covers the pre-flight that prices a draft batch before you submit it, including how many of its objects were already extracted and what a force or replace re-run would re-spend on GPU work; the deduplication default, which is keyed on the bucket and the collection together so re-triggering after a bucket grows only pays for what is new; a continuous bucket sync that polls the storage connection and submits what lands; and moving a collection nobody queries to cold, which takes its vectors out of the vector store while leaving them searchable. It also says plainly which of those the Python client can do and which are HTTP calls today.
- Sep 16, 2026The solution finder points at the right page, and every answer is now clickableThree questions on the finder linked to a page that did not answer them. Asking how to filter NSFW, violent or policy-violating content offered a classification recipe under a description of a moderation recipe, and the moderation recipe existed the whole time. Asking how to remove duplicates offered the same search page twice, once as the primary answer and once as the alternative. Asking how to jump to a moment in a video offered outlier detection. All three now go where the page's own description says they go. Separately, twelve of the recommended answers were University modules and the finder had no link for that kind of answer at all, so they rendered as text with nothing to click, including the top answer for four questions. They link now, and the page finally publishes its nineteen questions and answers as structured data an answer engine can read.
- Sep 15, 2026Every recipe page shows Python and curl that run, and retrieval steps that existAll 49 recipes now carry a Python snippet built on calls the SDK and API accept: a bucket with a schema, a collection on a real extractor, an upload with a blobs list, a trigger, and a retriever made of stages that exist, executed with inputs. Clustering, anomaly detection and drift recipes create clusters and read their groups, run metrics or CSV export. The taxonomy recipes build flat and hierarchical taxonomies, the bring-your-own-vector recipes search a standalone namespace by vector name, and the RAG recipes rerank passages and write the cited answer inside the retriever. Every request body in the snippets was checked against the API's request models, including extractor and stage parameters. Each retrieval step names a stage from the retriever catalog and links to it, extractor chips point at extractor pages, and the hand-typed run counts that read like usage numbers are gone from recipe and extractor cards. The problem finder at mixpeek.com/find-solution now links only to recipes that exist.