Mixedbread

Generated Metadata

Learn about the metadata generated by Mixedbread Stores for ingestion and search.

Mixedbread Stores automatically generate metadata for each file ingested. This generated metadata provides structured information about the content of the file, including language, size, headings, number of pages, and more.

The generated metadata can be retrieved using the generated_metadata chunk field.

Metadata Types

The generated_metadata object is a typed structure discriminated by the type field. If type is not present, it is inferred from file_type (MIME); otherwise it defaults to text.

Supported type values: markdown, text, email, csv, pdf, code, audio, video, image.

Every metadata object includes a file_extension field (e.g., ".mp3", ".pdf") reflecting the original file extension.

For a full list of supported file formats, see Supported File Types.

Markdown - Heading Extraction

When processing markdown files, the system automatically extracts and preserves heading structure to enhance search relevance and provide context. This feature works for all markdown formats (.md, .markdown, .mdx).

What You Get

Each markdown chunk’s generated_metadata includes:

  1. type: Always "markdown".
  2. file_type: Always "text/markdown".
  3. language: Detected language of the text.
  4. word_count: Word count for the chunk.
  5. file_size: File size in bytes.
  6. chunk_headings: Headings found within the current chunk ([{ level: number, text: string }]).
  7. heading_context: The document structure context leading up to this chunk ([{ level: number, text: string }]).
  8. start_line: Starting line number of the chunk in the source file.
  9. num_lines: Number of lines in the chunk.

Example Output

Consider this markdown document:

# Getting Started
## Installation
### Prerequisites
You need Python 3.8+ installed.
...

### Setup
Run the following command:

```bash
pip install package
```
...

## Configuration
### Environment Variables
Set these variables in your `.env` file.

### Database Setup
Configure your database connection.

When processed, the chunks would have generated_metadata like this:

Chunk 1 (Prerequisites section):

{
  "type": "markdown",
  "file_type": "text/markdown",
  "language": "en",
  "word_count": 6,
  "file_size": 1234,
  "chunk_headings": [
    {"level": 1, "text": "Getting Started"},
    {"level": 2, "text": "Installation"},
    {"level": 3, "text": "Prerequisites"}
  ],
  "heading_context": []
}

Chunk 2 (Setup section):

{
  "type": "markdown",
  "file_type": "text/markdown",
  "language": "en",
  "word_count": 12,
  "file_size": 1234,
  "chunk_headings": [
    {"level": 3, "text": "Setup"}
  ],
  "heading_context": [
    {"level": 1, "text": "Getting Started"},
    {"level": 2, "text": "Installation"},
    {"level": 3, "text": "Prerequisites"}
  ]
}

Chunk 3 (Configuration section):

{
  "type": "markdown",
  "file_type": "text/markdown",
  "language": "en",
  "word_count": 20,
  "file_size": 1234,
  "chunk_headings": [
    {"level": 2, "text": "Configuration"},
    {"level": 3, "text": "Environment Variables"},
    {"level": 3, "text": "Database Setup"}
  ],
  "heading_context": [
    {"level": 1, "text": "Getting Started"},
    {"level": 2, "text": "Installation"},
    {"level": 3, "text": "Setup"}
  ]
}

Text - Common Fields

Plain text chunks include a simpler generated_metadata shape:

  • type: "text"
  • file_type: "text/plain"
  • language: Detected language of the text
  • word_count: Word count for the chunk
  • file_size: File size in bytes
  • start_line: Starting line number of the chunk in the source file
  • num_lines: Number of lines in the chunk

Example

{
  "type": "text",
  "file_type": "text/plain",
  "language": "en",
  "word_count": 57,
  "file_size": 2048,
  "start_line": 0,
  "num_lines": 15
}

Email - Headers

Email files (.eml and Outlook .msg) carry their headers on every chunk, so a chunk deep in a long email can still be filtered by sender, recipient or date. Alongside the common text fields:

  • type: "email"
  • file_type: "message/rfc822" or "application/vnd.ms-outlook"
  • subject: The email subject
  • from_email: The sender's address
  • to_emails, cc_emails, bcc_emails: Recipient addresses
  • date: The sent date in ISO 8601, in UTC
  • attachment_names: File names of the attachments

Example

{
  "type": "email",
  "file_type": "message/rfc822",
  "language": "en",
  "word_count": 182,
  "file_size": 1210,
  "start_line": 0,
  "num_lines": 14,
  "subject": "Quarterly update",
  "from_email": "jane@example.com",
  "to_emails": ["team@example.com"],
  "cc_emails": [],
  "bcc_emails": [],
  "date": "2026-07-01T09:15:00+00:00",
  "attachment_names": ["report.pdf"]
}

CSV - Row Positions

CSV files are rendered as markdown tables. Each chunk repeats the header row and holds as many data rows as fit the chunk's token budget, so a chunk always reads as a complete table. A row is only split when it cannot fit a chunk on its own: it is then broken into column groups, each led by the row's first column, and a single oversized cell is broken into pieces. Alongside the common text fields, each chunk records where its rows live in the source file. Rows and lines are 1-based, and the header is row 1. An Excel sep= line or a lone title above the header is skipped but still counted. Tab-separated .tsv files use the same parser and metadata.

  • type: "csv"
  • file_type: "text/csv"
  • start_line: Line of the chunk's first data row in the source file
  • num_lines: Number of source lines the chunk's rows span (a quoted cell may span several lines)
  • start_row: Row number of the chunk's first data row
  • num_rows: Number of rows the chunk spans

The chunk's offset field is the character offset of the first data row.

Example

{
  "type": "csv",
  "file_type": "text/csv",
  "language": "en",
  "word_count": 1240,
  "file_size": 18320,
  "start_line": 42,
  "num_lines": 40,
  "start_row": 42,
  "num_rows": 40
}

Code - Language and Size

For supported source files (e.g., Python, TypeScript, Java, C#), generated_metadata includes:

  • type: "code"
  • file_type: One of text/x-python, text/x-script.python, application/typescript, text/typescript, text/x-java-source, text/x-csharp, or application/javascript
  • language: Detected programming language
  • word_count: Tokenized word count approximation for code
  • file_size: File size in bytes
  • start_line: Starting line number of the chunk in the source file
  • num_lines: Number of lines in the chunk

Example

{
  "type": "code",
  "file_type": "text/x-python",
  "language": "python",
  "word_count": 120,
  "file_size": 8192,
  "start_line": 0,
  "num_lines": 42
}

PDF - Document Stats

PDF chunks have specialized document-level stats:

  • type: "pdf"
  • file_type: "application/pdf"
  • total_pages: Total number of pages in the document
  • total_size: Total size of the original file in bytes

Example

{
  "type": "pdf",
  "file_type": "application/pdf",
  "total_pages": 42,
  "total_size": 1048576
}

Image - Dimensions

Image chunks include basic dimensional metadata:

  • type: "image"
  • file_type: String (MIME type, e.g., image/jpeg, image/png)
  • file_size: File size in bytes
  • width: Image width in pixels
  • height: Image height in pixels

Example

{
  "type": "image",
  "file_type": "image/jpeg",
  "file_size": 204800,
  "width": 1920,
  "height": 1080
}

Audio - Media Information

Audio chunks include specialized media metadata:

  • type: "audio"
  • file_type: String (MIME type, e.g., audio/mpeg, audio/wav)
  • file_size: File size in bytes
  • total_duration_seconds: Total duration of the audio in seconds
  • sample_rate: Audio sample rate in Hz
  • channels: Number of audio channels
  • audio_format: Audio format code
  • bpm: Detected beats per minute (optional, only present when detected)
  • speakers: Labels of the speakers heard in the chunk, such as ["Speaker 1", "Speaker 3"] (optional, see Speakers)
  • speaker_count: Number of speakers in the whole file (optional)
  • speaker_turns: The chunk's transcription as {"speaker", "text", "start_time_seconds", "end_time_seconds"} turns, timed in seconds from the start of the file (optional)
  • start_time_seconds: Start time of the chunk in seconds
  • end_time_seconds: End time of the chunk in seconds
  • duration_seconds: Duration of the chunk in seconds
  • chunk_size_bytes: Size of the chunk in bytes

Example

{
  "type": "audio",
  "file_type": "audio/mpeg",
  "file_size": 5242880,
  "total_duration_seconds": 180.5,
  "sample_rate": 44100,
  "channels": 2,
  "audio_format": 1,
  "speakers": ["Speaker 1", "Speaker 2"],
  "speaker_count": 4,
  "speaker_turns": [
    {
      "speaker": "Speaker 1",
      "text": "Welcome back to the show.",
      "start_time_seconds": 0.12,
      "end_time_seconds": 1.84
    },
    {
      "speaker": "Speaker 2",
      "text": "Thanks for having me.",
      "start_time_seconds": 2.3,
      "end_time_seconds": 3.51
    }
  ],
  "start_time_seconds": 0,
  "end_time_seconds": 38.93,
  "duration_seconds": 38.93,
  "chunk_size_bytes": 650736
}

Speakers

With the high_quality parsing strategy, an audio file or a video with an audio track is transcribed as a whole, and every word is attributed to a speaker. Speakers are numbered in the order they first speak, and a label means the same person in every chunk of the file, so you can filter on generated_metadata.speakers to find the chunks where one speaker talks. Each turn in speaker_turns runs from the start of its first word to the end of its last, so you can cite the moment a passage was said. Files longer than 4 hours or of 2 GB or more are transcribed one chunk at a time without speaker fields.

Video - Media Information

Video chunks include specialized media metadata:

  • type: "video"
  • file_type: String (MIME type, e.g., video/mp4, video/webm)
  • file_size: File size in bytes
  • total_duration_seconds: Total duration of the video in seconds
  • fps: Frames per second
  • width: Video width in pixels
  • height: Video height in pixels
  • frame_count: Total number of frames
  • has_audio_stream: Whether the video contains an audio track
  • bpm: Detected beats per minute (optional, only present when detected)
  • speakers: Labels of the speakers heard in the chunk, such as ["Speaker 1", "Speaker 3"] (optional, see Speakers)
  • speaker_count: Number of speakers in the whole file (optional)
  • speaker_turns: The chunk's transcription as {"speaker", "text", "start_time_seconds", "end_time_seconds"} turns, timed in seconds from the start of the file (optional)
  • start_time_seconds: Start time of the chunk in seconds
  • end_time_seconds: End time of the chunk in seconds
  • duration_seconds: Duration of the chunk in seconds
  • chunk_size_bytes: Size of the chunk in bytes

Example

{
  "type": "video",
  "file_type": "video/mp4",
  "file_size": 15728640,
  "total_duration_seconds": 120.0,
  "fps": 30.0,
  "width": 1920,
  "height": 1080,
  "frame_count": 3600,
  "has_audio_stream": true,
  "start_time_seconds": 0.0,
  "end_time_seconds": 120.0,
  "duration_seconds": 120.0,
  "chunk_size_bytes": 2097152
}

Type Inference and Flexibility

  • If type is not present in generated_metadata, it is inferred from file_type when possible.
  • If neither type nor a recognized file_type is present, the type defaults to "text".
  • The system may include additional fields as needed for future enhancements; clients should read the documented fields and ignore unknown ones.

On this page

Metadata TypesMarkdown - Heading ExtractionWhat You GetExample OutputText - Common FieldsExampleEmail - HeadersExampleCSV - Row PositionsExampleCode - Language and SizeExamplePDF - Document StatsExampleImage - DimensionsExampleAudio - Media InformationExampleSpeakersVideo - Media InformationExampleType Inference and Flexibility