Any file type.
One binary. Auto-detected.
CSV, images, video, PDF, audio, archives. The agent reads the content itself, figures out what it is, and picks the right way to process it. No format-specific extractors, no per-file config. One binary handles everything you throw at it.
Detected media
formats
The agent uses 10 format extractors to scan and classify every file by its actual content, not by extension.
| File Type | Auto-Detects | Chunking Strategy | Use Case |
|---|---|---|---|
| CSV / JSON / Parquet | Rows, columns, data types, headers | row_range | Tabular data, analytics pipelines, database exports |
| Images (JPEG, PNG, WebP, GIF) | Format, dimensions, color depth, EXIF, face count | file_per_chunk | Image classification, object detection, metadata extraction |
| Video (MP4, MOV, AVI, MKV) | Container, codec, duration, resolution, keyframes | byte_range / file_per_chunk | Video transcoding, keyframe extraction, content moderation |
| Page count, text content, embedded images, OCR text | file_per_chunk | Document processing, text extraction, form parsing | |
| Audio (MP3, WAV, FLAC) | Codec, sample rate, duration, channels | file_per_chunk | Speech-to-text, audio classification, transcription pipelines |
| Archives (ZIP, TAR, GZ, 7Z) | File listing, compression ratio, nested structure | extract + recurse | Batch processing, compressed data extraction, nested archives |
How auto-detection
works
The agent doesn't guess file types from extensions. It reads the actual content.
Content header scan
When a file is added to a dataset, the agent reads the first few bytes to identify the file format, not the file extension. A .csv file mislabeled as .txt is still detected as CSV.
Metadata extraction
The format-specific extractor pulls relevant metadata: dimensions for images, codec and duration for video, page count for PDF, sample rate for audio, compression ratio for archives.
Chunking strategy selection
Based on format type and metadata, the agent selects the optimal chunking strategy. row_range for tabular, file_per_chunk for images/PDF/audio, byte_range for video, extract+recurse for archives.
Plugin execution
Each chunk is dispatched to the selected plugin via the embedded sandbox. The plugin receives the chunk data with format metadata. No format-specific config needed on the plugin side.
Mixed-format
pipelines
One agent handles every format in a single pipeline. No separate infrastructure per file type.
E-commerce catalog enrichment
Input: CSV product listings + product images + PDF spec sheets. Pipeline: extract CSV rows → resize images → extract PDF specs → merge enriched catalog. One agent, three formats, one pipeline.
Medical imaging preprocessing
Input: DICOM-style images + CSV metadata + PDF reports. Pipeline: extract image features → parse metadata → OCR reports → combine into structured output. All processed locally. Zero egress.
Video content moderation
Input: MP4 video files + JSON annotations. Pipeline: extract keyframes → run classification on frames → transcribe audio → merge results. One agent handles video, audio, and structured data.
One binary.
Every format.
CSV, images, video, PDF, audio. Everything runs through the same agent, the same sandbox, the same pipeline. No format-specific infrastructure to buy or maintain.