Multimodal RAG

from openai import OpenAI

from mixpeek import Mixpeek

client = Mixpeek(api_key="YOUR_API_KEY", namespace="launch-recordings")
openai = OpenAI(api_key="YOUR_OPENAI_KEY")

# 1. Recordings are cut at pauses, transcribed, and each passage is embedded
bucket = client.buckets.create(
    bucket_name="launch-recordings",
    bucket_schema={"properties": {"recording": {"type": "video"}}},
)
collection = client.collections.create(
    collection_name="launch-recordings",
    source={"type": "bucket", "bucket_ids": [bucket["bucket_id"]]},
    feature_extractor={
        "feature_extractor_name": "multimodal_extractor",
        "version": "v1",
        "parameters": {
            "split_method": "silence",
            "run_transcription": True,
            "run_transcription_embedding": True,
        },
    },
)
client.buckets.upload(
    bucket["bucket_id"],
    blobs=[{"property": "recording", "type": "video", "data": "s3://your-bucket/launch/customer-call.mp4"}],
)
client.collections.trigger(collection["collection_id"])

# 2. Retrieve passages, then rerank them against the question
retriever = client.retrievers.create(
    retriever_name="launch-context",
    collection_identifiers=["launch-recordings"],
    input_schema={"query": {"type": "text", "required": True}},
    stages=[
        {
            "stage_name": "search",
            "stage_id": "feature_search",
            "parameters": {
                "searches": [
                    {
                        "feature_uri": "mixpeek://multimodal_extractor@v1/multilingual_e5_large_instruct_v1",
                        "query": {"input_mode": "text", "value": "{{INPUT.query}}"},
                        "top_k": 50,
                    },
                ],
                "final_top_k": 50,
            },
        },
        {
            "stage_name": "rerank",
            "stage_id": "rerank",
            "parameters": {
                "inference_name": "BAAI__bge_reranker_v2_m3",
                "query": "{{INPUT.query}}",
                "document_field": "transcription",
                "top_k": 8,
            },
        },
    ],
)

question = "How did the product launch go?"
results = client.retrievers.execute(retriever["retriever_id"], inputs={"query": question})

# 3. Number each passage with its source and timestamp so the answer can cite it
context = " ".join(
    f"[{i}] {doc['transcription']} (source {doc['root_object_id']} at {doc['start_time']}s)."
    for i, doc in enumerate(results["documents"], 1)
)
response = openai.chat.completions.create(
    model="gpt-4o",
    messages=[
        {"role": "system", "content": "Answer from the numbered context and cite the numbers. Context: " + context},
        {"role": "user", "content": question},
    ],
)
print(response.choices[0].message.content)