Retrieval & Document Workflows

Bedrock Knowledge Bases Multimodal Chunking: Audio, Video, and the Data Automation Parser

Your searchable knowledge is not all PDFs. Recorded trainings, support calls, and all-hands videos can feed a Bedrock Knowledge Base too — if you understand the two paths media takes on its way into the index.

Claude 3P 101 · Updated July 2026 · Unofficial guide

Amazon Bedrock Knowledge Bases is best known for text RAG: chunk documents, embed the chunks, retrieve them at query time, and have a generation model — Claude is explicitly supported for this — answer with citations. Less well known is that the ingestion side also handles audio and video. There are two distinct mechanisms, and choosing between them shapes both retrieval quality and what your citations can point at.

Path 1: chunk the media itself, with Nova multimodal embeddings

When a knowledge base uses Nova multimodal embeddings — one of the documented 1024-dimension multimodal options alongside Titan Multimodal Embeddings G1 and Cohere Embed v3 multimodal — audio and video files can be chunked as media. Instead of a chunk being roughly 300 tokens of text, a chunk is a slice of time: the chunk duration is configurable from 1 to 30 seconds, with a 5-second default. Each time-slice is embedded directly into the same vector space as your text, so a query like "the moment the presenter demos the failover" can retrieve the actual segment of the recording rather than a description of it.

Duration tuning follows the same logic as text chunk sizing. Short chunks (a few seconds) give precise retrieval — good for dense, fast-moving content like product demos — but fragment context. Longer chunks (up to 30 seconds) keep a full spoken thought together — better for lectures and meetings — at the cost of retrieving more irrelevant material around each hit. The 5-second default is a starting point, not a recommendation for every corpus; evaluate against real queries before scaling ingestion, since re-chunking means re-embedding the library.

Path 2: convert media to text first, with the Data Automation parser

The alternative route keeps the vector index purely textual. The Bedrock Data Automation parser converts media files into text — transcripts for speech and scene summaries for visual content — and that generated text then flows through the standard text chunking pipeline (default, fixed-size, hierarchical, or semantic; see the chunking strategies article). From the retrieval system's point of view, your webinar has become a document.

This path is the natural fit when the value of the media is what was said: recorded meetings, support calls, training sessions, town halls. It also means your existing text embedding model (Titan or Cohere families) covers everything — no multimodal embedding model required — and retrieved chunks are human-readable text, which makes answers easy to audit. What you give up is anything the transcript and scene summaries fail to capture: visual detail, tone, and exact timing live only in the original file.

Media chunking (Nova multimodal)Data Automation parser
Chunk unit1–30 s media segment (default 5 s)Text chunk of transcript / scene summary
Embedding modelNova multimodal embeddingsStandard text embeddings
RetrievesThe media segment itselfText describing the media
Best forVisually meaningful contentSpeech-heavy recordings

What this pipeline is designed for — and what it isn't

The multimodal ingestion path targets organizations whose institutional knowledge is locked in recordings and mixed-media documents: onboarding videos, compliance trainings, customer calls, conference talks. It is an ingestion feature — it gets media into a searchable index. It is not a promise that the generation model watches video at answer time: with the parser path, Claude reasons over transcript text; for document parsing more broadly, Claude vision models are explicitly listed as usable parsers for advanced document parsing (alongside Nova vision and Llama 4 vision), which covers visually complex documents rather than long-form video.

Rule of thumb: if a human would answer the query by reading a transcript, use the Data Automation parser path. If they would need to see or hear the moment itself, chunk the media with Nova multimodal embeddings.

Two operational reminders regardless of path: chunking parameters are fixed at ingestion, so pilot on a representative slice of your media library first; and this is all Bedrock-native machinery — it runs on Bedrock's agent-runtime APIs, not the Anthropic Messages API, so none of it transfers to Vertex AI or Foundry, where the documented Claude pattern is bring-your-own-retrieval.

Where to go next

For the architectural decision above this one, read managed vs customer-managed knowledge bases; for the embedding-model options, see Knowledge Base embeddings; and for querying with Claude as the generator, RetrieveAndGenerate.

Sources