Joins
Multimodal RAG (Evidence-Grounded Generation)
Retrieval + grouping + LLM generation. The recipe is the retriever config—not the LLM. Enables citations back to source timestamps.
video
image
text
audio
Multi-Stage
67.0K runs
Deploy RecipeWhy This Matters
RAG is just retrieval + external LLM. Mixpeek handles the retrieval infrastructure, you bring the generation model.
from mixpeek import Mixpeekfrom openai import OpenAImixpeek = Mixpeek(api_key="your-mixpeek-key")openai = OpenAI(api_key="your-openai-key")# Retrieve context with citationscontext = mixpeek.retrievers.execute(retriever_id="rag-retriever",inputs={"query_text": "How did the product launch go?","return_citations": True},limit=5)# Format context with sourcescontext_str = "\n".join([f"[{i+1}] {doc['text']} (Source: {doc['source_url']} @ {doc['timestamp']})"for i, doc in enumerate(context['documents'])])# Generate with LLMresponse = openai.chat.completions.create(model="gpt-4",messages=[{"role": "system", "content": f"Context:\n{context_str}"},{"role": "user", "content": "Summarize the product launch feedback"}])
Retrieval Flow
1
feature search(search)
Semantic search for relevant context
2
llm rerank(rank)
Rerank by relevance to query
3
limit(reduce)
Top-k most relevant chunks
Feature Extractors
Feature Extractors
Text Embedding
Extract semantic embeddings from documents, transcripts and text content
827K runs
Image Embedding
Generate visual embeddings for similarity search and clustering
752K runs
Video Embedding
Generate vector embeddings for video content
610K runs
Retriever Stages
feature search
Search collections using multimodal embeddings
search
llm rerank
Rerank documents using LLM-based relevance scoring
rank
limit
Limit the number of documents returned
reduce
