OCR integration

Added in version 2.3.

To deal with images containing text, just install Tesseract. Tesseract will be auto-detected by Tika or you can explicitly set the path to tesseract binary. Then add an image (png, jpg, …) into your FSCrawler Root directory. After the next index update, the text will be indexed and placed in _source.content.

OCR settings

Here is a list of OCR settings (under fs.ocr prefix):

Name	Environment Variable	Default value	Documentation
`fs.ocr.enabled`	`FSCRAWLER_FS_OCR_ENABLED`	`true`	Disable/Enable OCR
`fs.ocr.language`	`FSCRAWLER_FS_OCR_LANGUAGE`	`"eng"`	OCR Language
`fs.ocr.path`	`FSCRAWLER_FS_OCR_PATH`	`null`	OCR Path
`fs.ocr.data_path`	`FSCRAWLER_FS_OCR_DATA_PATH`	`null`	OCR Data Path
`fs.ocr.output_type`	`FSCRAWLER_FS_OCR_OUTPUT_TYPE`	`txt`	OCR Output Type
`fs.ocr.pdf_strategy`	`FSCRAWLER_FS_OCR_PDF_STRATEGY`	`auto`	OCR PDF Strategy
`fs.ocr.page_seg_mode`	`FSCRAWLER_FS_OCR_PAGE_SEG_MODE`	`01`	OCR Page Seg Mode
`fs.ocr.preserve_interword_spacing`	`FSCRAWLER_FS_OCR_PRESERVE_INTERWORD_SPACING`	`false`	OCR Preserve Interword Spacing

Disable/Enable OCR

Added in version 2.7.

You can completely disable using OCR by setting fs.ocr.enabled property in your ~/.fscrawler/test/_settings.yaml file:

name: "test"
fs:
  url: "/path/to/data/dir"
  ocr:
    enabled: false

By default, OCR is activated if tesseract can be found on your system.

OCR Language

If you are using the default Docker image (see Using Docker) or if you have installed any of the Tesseract Languages, you can use them when parsing your documents by setting fs.ocr.language property in your ~/.fscrawler/test/_settings.yaml file:

name: "test"
fs:
  url: "/path/to/data/dir"
  ocr:
    language: "eng"

Note

You can define multiple languages by using + sign as a separator:

name: "test"
fs:
  url: "/path/to/data/dir"
  ocr:
    language: "eng+fas+fra"

OCR Path

If your Tesseract application is not available in default system PATH, you can define the path to use by setting fs.ocr.path property in your ~/.fscrawler/test/_settings.yaml file:

name: "test"
fs:
  url: "/path/to/data/dir"
  ocr:
    path: "/path/to/tesseract/bin/"

You can point fs.ocr.path either at the tesseract executable itself or at the directory that contains it: FSCrawler accepts both forms.

When you set it, it’s highly recommended to set the OCR Data Path.

OCR Data Path

Set the path to the ‘tessdata’ folder, which contains language files and config files if Tesseract can not be automatically detected. You can define the path to use by setting fs.ocr.data_path property in your ~/.fscrawler/test/_settings.yaml file:

name: "test"
fs:
  url: "/path/to/data/dir"
  ocr:
    path: "/path/to/tesseract/bin/"
    data_path: "/path/to/tesseract/share/tessdata/"

OCR Output Type

Added in version 2.5.

Set the output type from ocr process. fs.ocr.output_type property can be defined to txt or hocr in your ~/.fscrawler/test/_settings.yaml file:

name: "test"
fs:
  url: "/path/to/data/dir"
  ocr:
    output_type: "hocr"

Note

When omitted, txt value is used.

OCR PDF Strategy

By default, FSCrawler will also try to extract also images from your PDF documents and run OCR on them. This can be a CPU intensive operation. If you don’t mean to run OCR on PDF but only on images, you can set fs.ocr.pdf_strategy to "no_ocr" or to "auto":

name: "test"
fs:
  ocr:
    pdf_strategy: "auto"

Supported strategies are:

auto: No OCR is performed on PDF documents if there is more than 10 characters extracted. See PDFParser OCR Options.
no_ocr: No OCR is performed on PDF documents. OCR might be performed on images though if OCR is not disabled. See Disable/Enable OCR.
ocr_only: Only OCR is performed.
ocr_and_text: OCR and text extraction is performed.

Note

When omitted, auto value is used. OCR is skipped on PDF pages that already contain more than 10 characters of text. If you need OCR on every page regardless, set ocr_and_text or ocr_only.

OCR Page Seg Mode

Set Tesseract to only run a subset of layout analysis and assume a certain form of image. The options for N are:

0 = Orientation and script detection (OSD) only.
1 = Automatic page segmentation with OSD.
2 = Automatic page segmentation, but no OSD, or OCR. (not implemented)
3 = Fully automatic page segmentation, but no OSD.
4 = Assume a single column of text of variable sizes.
5 = Assume a single uniform block of vertically aligned text.
6 = Assume a single uniform block of text.
7 = Treat the image as a single text line.
8 = Treat the image as a single word.
9 = Treat the image as a single word in a circle.
10 = Treat the image as a single character.
11 = Sparse text. Find as much text as possible in no particular order.
12 = Sparse text with OSD.
13 = Raw line. Treat the image as a single text line, bypassing hacks that are Tesseract-specific.

OCR Preserve Interword Spacing

Spaces between the words will be deleted.

Using a Vision Language Model (VLM) for OCR

Added in version 3.0.

This is an optional alternative to Tesseract. It is never activated unless you set fs.tika_config_path and configure a VLM parser in the JSON file. Without a custom Tika configuration, FSCrawler keeps using Tesseract for OCR (when available).

FSCrawler ships with Apache Tika’s tika-vlm module, which can send images — including PDF pages rendered by the PDF parser — to a Vision Language Model (VLM) instead of Tesseract. The module provides parsers for OpenAI-compatible chat completions endpoints (openai-vlm-parser and openai-vlm-deterministic-parser, which work with vLLM, Ollama, OpenRouter, Azure OpenAI, LiteLLM…), Anthropic Claude (claude-vlm-parser) and Google Gemini (gemini-vlm-parser).

Choosing a parser

Component	When to use it
(none — Tesseract)	Default FSCrawler behaviour: no `fs.tika_config_path`, or a custom config that does not enable a VLM parser. Tesseract handles OCR when installed.
`openai-vlm-deterministic-parser`	Recommended for OCR via an OpenAI-compatible VLM endpoint. Same JSON settings as `openai-vlm-parser`, but forces `temperature: 0` (greedy decoding) to avoid hallucinations on small vision models.
`openai-vlm-parser`	OCR via an OpenAI-compatible endpoint without forcing the sampling temperature. Prefer this only when you explicitly need the server’s default sampling behaviour.
`claude-vlm-parser`	OCR delegated to an Anthropic Claude vision endpoint.
`gemini-vlm-parser`	OCR delegated to a Google Gemini vision endpoint.

The VLM parsers are configured through a custom Tika configuration file (see Local FS settings, fs.tika_config_path). The example below routes OCR to a local vLLM server running an OpenAI-compatible endpoint on http://localhost:8000:

{
  "parsers": [
    { "default-parser": { "exclude": ["tesseract-ocr-parser"] } },
    { "pdf-parser": {
        "ocr": {
          "strategy": "AUTO",
          "maxPagesToOcr": 10
        }
      } },
    { "openai-vlm-deterministic-parser": {
        "baseUrl": "http://localhost:8000",
        "model": "Qwen/Qwen2.5-VL-7B-Instruct",
        "maxTokens": 4096,
        "timeoutSeconds": 300
      } }
  ]
}

name: "test"
fs:
  tika_config_path: '/path/to/tikaConfig.json'

Some notes about this configuration:

A VLM parser must be listed explicitly in the JSON configuration to be used for OCR.
If your configuration uses default-parser, exclude the VLM parser components you do not need (see the example in Local FS settings). Otherwise Tika may load them through SPI when tika-vlm is on the classpath.
Without a custom Tika configuration, FSCrawler’s built-in parser chain keeps the VLM parsers disabled and continues to use Tesseract.
default-parser with exclude removes Tesseract from the parser chain, so OCR-able images are claimed by the VLM parser instead. If you leave Tesseract in, it takes precedence when installed.
The PDF parser renders pages according to the ocr.strategy and hands them to the OCR parser — the VLM in this setup. Always set maxPagesToOcr explicitly: every OCR’ed page is one VLM request, so an unbounded value can be slow and expensive on large documents.
apiKey can be set for hosted endpoints; for a local vLLM server it can usually be omitted.
Extracted metadata includes the model used and token usage (vlm:model, vlm:prompt_tokens, vlm:completion_tokens) in the raw metadata.

Deterministic decoding (`openai-vlm-deterministic-parser`)

Tika’s openai-vlm-parser does not send a temperature field, so the server decides the sampling temperature (vLLM typically defaults to 1.0). Small OCR-oriented vision models hallucinate heavily under sampling — measured against a local PaddleOCR-VL-1.6-0.9B, the very same request returned the actual page text with temperature: 0 and Chinese hallucinations with <|LOC_*|> layout tokens without it. FSCrawler therefore ships openai-vlm-deterministic-parser, a drop-in variant of openai-vlm-parser that forces "temperature": 0 (greedy decoding) on every request. It accepts exactly the same configuration keys; simply use it in place of openai-vlm-parser in the custom Tika configuration.

Warning

The VLM parser performs a single health check (GET {baseUrl}/v1/models) when it is initialized, at crawler startup. If the VLM server is not reachable at that moment, the parser marks itself unavailable and silently skips OCR for the whole lifetime of the crawler — no per-document retry is attempted. Make sure the VLM server is up before starting FSCrawler, and restart FSCrawler if the server was down when it started.

OCR integration

OCR settings

Disable/Enable OCR

OCR Language

OCR Path

OCR Data Path

OCR Output Type

OCR PDF Strategy

OCR Page Seg Mode

OCR Preserve Interword Spacing

Using a Vision Language Model (VLM) for OCR

Choosing a parser

Deterministic decoding (openai-vlm-deterministic-parser)

Deterministic decoding (`openai-vlm-deterministic-parser`)