diff --git a/litellm-rust/src/llms/base_llm/ocr/transformation.rs b/litellm-rust/src/llms/base_llm/ocr/transformation.rs
new file mode 100644
index 00000000000..6bd9ac17b51
--- /dev/null
+++ b/litellm-rust/src/llms/base_llm/ocr/transformation.rs
@@ -0,0 +1,146 @@
+//! Base OCR transformation: pure serde types and the `BaseOcrConfig` trait.
+//!
+//! Mirrors `litellm/llms/base_llm/ocr/transformation.py`. No PyO3, no I/O.
+
+use serde::{Deserialize, Serialize};
+use serde_json::{Map, Value};
+
+/// Document passed to a provider. Always a dict with `type="document_url"`
+/// or `type="image_url"` by the time it reaches a provider transform.
+///
+/// We model it as a free-form JSON object so extra keys are preserved and
+/// the document is passed through to the upstream body verbatim.
+pub type DocumentType = Value;
+
+/// HTTP header pairs (name, value).
+pub type Headers = Vec<(String, String)>;
+
+/// Page dimensions from an OCR response.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub struct OcrPageDimensions {
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub dpi: Option<i64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub height: Option<i64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub width: Option<i64>,
+}
+
+/// Image extracted from an OCR page. Extra keys are preserved (`extra = "allow"`).
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub struct OcrPageImage {
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub image_base64: Option<String>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub bbox: Option<Value>,
+
+    /// Preserve any additional fields returned by the provider.
+    #[serde(flatten)]
+    pub extra: Map<String, Value>,
+}
+
+/// A single page from an OCR response. Extra keys are preserved.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub struct OcrPage {
+    pub index: i64,
+    pub markdown: String,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub images: Option<Vec<OcrPageImage>>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub dimensions: Option<OcrPageDimensions>,
+
+    /// Preserve any additional fields returned by the provider.
+    #[serde(flatten)]
+    pub extra: Map<String, Value>,
+}
+
+/// Usage information from an OCR response. Extra keys are preserved.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub struct OcrUsageInfo {
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub pages_processed: Option<i64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub doc_size_bytes: Option<i64>,
+
+    /// Preserve any additional fields returned by the provider.
+    #[serde(flatten)]
+    pub extra: Map<String, Value>,
+}
+
+/// Standard OCR response format (standardized to the Mistral OCR format).
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub struct OcrResponse {
+    pub pages: Vec<OcrPage>,
+    pub model: String,
+    #[serde(default)]
+    pub document_annotation: Option<Value>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub usage_info: Option<OcrUsageInfo>,
+    #[serde(default = "default_object")]
+    pub object: String,
+}
+
+fn default_object() -> String {
+    "ocr".to_string()
+}
+
+/// Inputs for a single OCR call, assembled by the caller (Python or HTTP server).
+#[derive(Debug, Clone)]
+pub struct OcrRequest {
+    /// Provider-stripped model id, used directly as body `"model"`.
+    pub model: String,
+    /// Pre-normalized document dict, passed through verbatim into the body.
+    pub document: DocumentType,
+    /// Optional explicit API key. If `None`, providers fall back to env.
+    pub api_key: Option<String>,
+    /// Optional API base override.
+    pub api_base: Option<String>,
+    /// Raw OCR params; providers filter to their supported list.
+    pub optional_params: Map<String, Value>,
+}
+
+/// Provider-agnostic OCR transform contract. Pure: no I/O.
+pub trait BaseOcrConfig {
+    /// Supported OCR parameter keys for this provider/model.
+    fn get_supported_ocr_params(&self, model: &str) -> Vec<String>;
+
+    /// Filter `optional_params` down to the supported list, dropping the rest.
+    fn map_ocr_params(
+        &self,
+        optional_params: &Map<String, Value>,
+        model: &str,
+    ) -> Map<String, Value> {
+        let supported = self.get_supported_ocr_params(model);
+        optional_params
+            .iter()
+            .filter(|(k, _)| supported.iter().any(|s| s == *k))
+            .map(|(k, v)| (k.clone(), v.clone()))
+            .collect()
+    }
+
+    /// Resolve the API key (param or env) and build request headers.
+    fn validate_environment(
+        &self,
+        model: &str,
+        api_key: Option<&str>,
+        env_lookup: &dyn Fn(&str) -> Option<String>,
+    ) -> Result<Headers, String>;
+
+    /// Build the complete OCR endpoint URL.
+    fn get_complete_url(&self, api_base: Option<&str>, model: &str) -> String;
+
+    /// Build the request body for the upstream OCR call.
+    fn transform_ocr_request(
+        &self,
+        model: &str,
+        document: &DocumentType,
+        optional_params: &Map<String, Value>,
+    ) -> Result<Value, String>;
+
+    /// Parse the upstream JSON body into the standard `OcrResponse`.
+    fn transform_ocr_response(
+        &self,
+        model: &str,
+        response_json: &Value,
+    ) -> Result<OcrResponse, String>;
+}