diff --git a/crates/artificial-core/Cargo.toml b/crates/artificial-core/Cargo.toml
index 116a98d..5003fff 100644
--- a/crates/artificial-core/Cargo.toml
+++ b/crates/artificial-core/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "artificial-core"
-version = "0.4.2"
+version = "0.5.0"
 edition = "2021"
 description = "Provider-agnostic core traits, generic client and error types for the Artificial prompt-engineering SDK"
 license = "MIT"
diff --git a/crates/artificial-core/src/client.rs b/crates/artificial-core/src/client.rs
index 655a163..14cee73 100644
--- a/crates/artificial-core/src/client.rs
+++ b/crates/artificial-core/src/client.rs
@@ -106,6 +106,8 @@ impl<B: ChatCompletionProvider> ChatCompletionProvider for ArtificialClient<B> {
 }
 
 impl<B: StreamingChatProvider> StreamingChatProvider for ArtificialClient<B> {
+    type Message = B::Message;
+
     type Delta<'s>
         = B::Delta<'s>
     where
diff --git a/crates/artificial-core/src/provider/chat_complete.rs b/crates/artificial-core/src/provider/chat_complete.rs
index 216a3c7..f48d3b2 100644
--- a/crates/artificial-core/src/provider/chat_complete.rs
+++ b/crates/artificial-core/src/provider/chat_complete.rs
@@ -39,7 +39,10 @@ pub trait ChatCompletionProvider: Send + Sync {
 /// Tool-call and richer payload support can be layered on later by
 /// introducing a dedicated enum – starting with plain text keeps the API
 /// minimal and backend-agnostic.
-pub trait StreamingChatProvider: ChatCompletionProvider {
+pub trait StreamingChatProvider: Send + Sync {
+    /// Chat message type consumed by this backend.
+    type Message: Send + Sync + 'static;
+
     /// The item type returned on the stream.  For now it is plain UTF-8 text
     /// chunks, but back-ends are free to wrap it in richer enums if needed.
     type Delta<'s>: Stream<Item = Result<String>> + Send + 's
diff --git a/crates/artificial-openai/Cargo.toml b/crates/artificial-openai/Cargo.toml
index 0fec55e..eb13812 100644
--- a/crates/artificial-openai/Cargo.toml
+++ b/crates/artificial-openai/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "artificial-openai"
-version = "0.4.2"
+version = "0.5.0"
 edition = "2024"
 description = "OpenAI backend adapter for the Artificial prompt-engineering SDK"
 license = "MIT"
@@ -24,7 +24,12 @@ reqwest = { version = "0.12", default-features = false, features = [
     "deflate",
     "multipart",
 ] }
-artificial-core = { path = "../artificial-core" , version = "0.4.2"}
+artificial-core = { path = "../artificial-core" , version = "0.5.0"}
 futures-util = "0.3"
 async-stream = "0.3"
 bytes = "1"
+tracing = { version = "0.1", optional = true }
+
+[features]
+default = []
+tracing = ["dep:tracing"]
diff --git a/crates/artificial-openai/src/adapter.rs b/crates/artificial-openai/src/adapter.rs
index 021038b..463152d 100644
--- a/crates/artificial-openai/src/adapter.rs
+++ b/crates/artificial-openai/src/adapter.rs
@@ -2,7 +2,7 @@ use std::{env, sync::Arc};
 
 use artificial_core::error::{ArtificialError, Result};
 
-use crate::client::OpenAiClient;
+use crate::client::{OpenAiClient, RetryPolicy};
 
 /// Thin wrapper that wires the HTTP client [`OpenAiClient`] into a value that
 /// implements [`artificial_core::backend::Backend`].
@@ -40,6 +40,7 @@ impl OpenAiAdapter {}
 #[derive(Default)]
 pub struct OpenAiAdapterBuilder {
     pub(crate) api_key: Option<String>,
+    pub(crate) retry: Option<RetryPolicy>,
 }
 
 impl OpenAiAdapterBuilder {
@@ -57,19 +58,34 @@ impl OpenAiAdapterBuilder {
     pub fn new_from_env() -> Self {
         Self {
             api_key: env::var("OPENAI_API_KEY").ok(),
+            retry: None,
         }
     }
 
+    /// Set a retry policy for OpenAI HTTP calls.
+    pub fn with_retry_policy(mut self, retry: RetryPolicy) -> Self {
+        self.retry = Some(retry);
+        self
+    }
+
     /// Finalise the builder and return a ready-to-use adapter.
     ///
     /// # Errors
     ///
     /// * [`ArtificialError::Invalid`] – if the API key is missing.
     pub fn build(self) -> Result<OpenAiAdapter> {
+        let api_key = self.api_key.ok_or(ArtificialError::Invalid(
+            "missing env variable: `OPENAI_API_KEY`".into(),
+        ))?;
+
+        let client = if let Some(retry) = self.retry {
+            OpenAiClient::new(api_key).with_retry_policy(retry)
+        } else {
+            OpenAiClient::new(api_key)
+        };
+
         Ok(OpenAiAdapter {
-            client: Arc::new(OpenAiClient::new(self.api_key.ok_or(
-                ArtificialError::Invalid("missing env variable: `OPENAI_API_KEY`".into()),
-            )?)),
+            client: Arc::new(client),
         })
     }
 }
diff --git a/crates/artificial-openai/src/client.rs b/crates/artificial-openai/src/client.rs
index eeb29cb..05372ac 100644
--- a/crates/artificial-openai/src/client.rs
+++ b/crates/artificial-openai/src/client.rs
@@ -10,9 +10,122 @@ use std::time::Duration;
 
 use crate::{
     api_v1::{ChatCompletionChunkResponse, ChatCompletionRequest, ChatCompletionResponse},
-    error::OpenAiError,
+    error::{OpenAiError, OpenAiRateLimitHeaders},
 };
 
+fn parse_retry_after_seconds(headers: &reqwest::header::HeaderMap) -> Duration {
+    use reqwest::header::RETRY_AFTER;
+    if let Some(val) = headers.get(RETRY_AFTER).and_then(|hv| hv.to_str().ok())
+        && let Ok(secs) = val.trim().parse::<u64>()
+    {
+        return Duration::from_secs(secs);
+    }
+    Duration::from_secs(0)
+}
+
+fn header_u32(headers: &reqwest::header::HeaderMap, name: &str) -> Option<u32> {
+    headers
+        .get(name)
+        .and_then(|hv| hv.to_str().ok())
+        .and_then(|s| s.parse::<u32>().ok())
+}
+
+fn header_string(headers: &reqwest::header::HeaderMap, name: &str) -> Option<String> {
+    headers
+        .get(name)
+        .and_then(|hv| hv.to_str().ok())
+        .map(|s| s.to_string())
+}
+
+fn extract_rate_limit_info(
+    headers: &reqwest::header::HeaderMap,
+) -> (Option<Duration>, Option<String>, OpenAiRateLimitHeaders) {
+    let retry_after = {
+        let d = parse_retry_after_seconds(headers);
+        if d.as_secs() > 0 { Some(d) } else { None }
+    };
+
+    let info = OpenAiRateLimitHeaders {
+        limit_requests: header_u32(headers, "x-ratelimit-limit-requests"),
+        remaining_requests: header_u32(headers, "x-ratelimit-remaining-requests"),
+        reset_requests: header_string(headers, "x-ratelimit-reset-requests"),
+        limit_tokens: header_u32(headers, "x-ratelimit-limit-tokens"),
+        remaining_tokens: header_u32(headers, "x-ratelimit-remaining-tokens"),
+        reset_tokens: header_string(headers, "x-ratelimit-reset-tokens"),
+    };
+
+    // Prefer request reset, fall back to token reset.
+    let reset_at = info
+        .reset_requests
+        .clone()
+        .or_else(|| info.reset_tokens.clone());
+
+    (retry_after, reset_at, info)
+}
+#[cfg(feature = "tracing")]
+fn log_rate_limit_tight(headers: &reqwest::header::HeaderMap, context: &str) {
+    let rem_reqs = header_u32(headers, "x-ratelimit-remaining-requests").unwrap_or(u32::MAX);
+    let rem_tokens = header_u32(headers, "x-ratelimit-remaining-tokens").unwrap_or(u32::MAX);
+    let lim_reqs = header_u32(headers, "x-ratelimit-limit-requests").unwrap_or(0);
+    let lim_tokens = header_u32(headers, "x-ratelimit-limit-tokens").unwrap_or(0);
+
+    // Heuristics: warn when headroom is tight
+    let tight_reqs = rem_reqs <= 2 || (lim_reqs > 0 && rem_reqs as f32 / lim_reqs as f32 <= 0.05);
+    let tight_tokens =
+        rem_tokens <= 128 || (lim_tokens > 0 && rem_tokens as f32 / lim_tokens as f32 <= 0.05);
+
+    if tight_reqs || tight_tokens {
+        tracing::warn!(
+            context,
+            remaining_requests = rem_reqs,
+            limit_requests = lim_reqs,
+            remaining_tokens = rem_tokens,
+            limit_tokens = lim_tokens,
+            "rate limit headroom is tight"
+        );
+    } else {
+        tracing::debug!(
+            context,
+            remaining_requests = rem_reqs,
+            limit_requests = lim_reqs,
+            remaining_tokens = rem_tokens,
+            limit_tokens = lim_tokens,
+            "rate limit status"
+        );
+    }
+}
+
+#[derive(Clone, Debug)]
+pub struct RetryPolicy {
+    pub max_retries: u32,
+    pub base_delay: Duration,
+    pub max_delay: Duration,
+    pub respect_retry_after: bool,
+}
+
+impl Default for RetryPolicy {
+    fn default() -> Self {
+        Self {
+            max_retries: 3,
+            base_delay: Duration::from_millis(500),
+            max_delay: Duration::from_secs(30),
+            respect_retry_after: true,
+        }
+    }
+}
+
+impl RetryPolicy {
+    fn backoff_for(&self, attempt: u32) -> Duration {
+        let pow = attempt.min(10);
+        let backoff = self.base_delay.saturating_mul(1 << pow);
+        if backoff > self.max_delay {
+            self.max_delay
+        } else {
+            backoff
+        }
+    }
+}
+
 const DEFAULT_BASE_URL: &str = "https://api.openai.com/v1";
 
 /// Minimal HTTP client for OpenAI’s *chat/completions* endpoint.
@@ -26,6 +139,7 @@ pub struct OpenAiClient {
     api_key: String,
     http: HttpClient,
     base: String,
+    retry: RetryPolicy,
 }
 
 impl OpenAiClient {
@@ -51,6 +165,123 @@ impl OpenAiClient {
             api_key: api_key.into(),
             http,
             base: base_url.unwrap_or_else(|| DEFAULT_BASE_URL.to_owned()),
+            retry: RetryPolicy::default(),
+        }
+    }
+
+    /// Allow callers to override the default retry policy.
+    pub fn with_retry_policy(mut self, retry: RetryPolicy) -> Self {
+        self.retry = retry;
+        self
+    }
+
+    // Internal: send POST with retry/backoff handling.
+    async fn post_json_with_retry(
+        &self,
+        url: String,
+        headers: HeaderMap,
+        request: &ChatCompletionRequest,
+    ) -> Result<reqwest::Response, OpenAiError> {
+        let mut attempt: u32 = 0;
+        loop {
+            let res = self
+                .http
+                .post(url.clone())
+                .headers(headers.clone())
+                .json(request)
+                .send()
+                .await;
+
+            match res {
+                Ok(resp) => {
+                    let status = resp.status();
+                    if status.is_success() {
+                        #[cfg(feature = "tracing")]
+                        {
+                            log_rate_limit_tight(resp.headers(), "success");
+                        }
+                        return Ok(resp);
+                    }
+
+                    let should_retry = status == reqwest::StatusCode::TOO_MANY_REQUESTS
+                        || status.is_server_error();
+
+                    if should_retry && attempt < self.retry.max_retries {
+                        let mut delay = self.retry.backoff_for(attempt);
+                        #[allow(unused_assignments)]
+                        let mut hdr_delay = Duration::from_secs(0);
+                        if self.retry.respect_retry_after {
+                            hdr_delay = parse_retry_after_seconds(resp.headers());
+                            if hdr_delay > delay {
+                                delay = hdr_delay;
+                            }
+                        }
+                        #[cfg(feature = "tracing")]
+                        {
+                            tracing::info!(
+                                attempt = attempt,
+                                status = %status,
+                                backoff_ms = delay.as_millis() as u64,
+                                retry_after_ms = hdr_delay.as_millis() as u64,
+                                "retrying request due to transient status"
+                            );
+                            log_rate_limit_tight(resp.headers(), "retrying");
+                        }
+                        // Blocking sleep to avoid introducing a new async runtime dependency.
+                        std::thread::sleep(delay);
+                        attempt += 1;
+                        continue;
+                    } else {
+                        let status = resp.status();
+                        let headers_map = resp.headers().clone();
+                        let body = resp.text().await.unwrap_or_default();
+                        if status == reqwest::StatusCode::TOO_MANY_REQUESTS {
+                            let (retry_after, reset_at, headers) =
+                                extract_rate_limit_info(&headers_map);
+                            #[cfg(feature = "tracing")]
+                            {
+                                let ra_ms = retry_after.map(|d| d.as_millis() as u64);
+                                tracing::warn!(
+                                    status = %status,
+                                    retry_after_ms = ?ra_ms,
+                                    reset_at = ?reset_at,
+                                    "rate limited; giving up after retries"
+                                );
+                            }
+                            return Err(OpenAiError::RateLimited {
+                                status,
+                                body,
+                                retry_after,
+                                reset_at,
+                                headers,
+                            });
+                        } else {
+                            return Err(OpenAiError::Api { status, body });
+                        }
+                    }
+                }
+                Err(err) => {
+                    // Retry on transport errors up to max_retries.
+                    if attempt < self.retry.max_retries
+                        && (err.is_timeout() || err.is_connect() || !err.is_status())
+                    {
+                        let delay = self.retry.backoff_for(attempt);
+                        #[cfg(feature = "tracing")]
+                        {
+                            tracing::info!(
+                                attempt = attempt,
+                                backoff_ms = delay.as_millis() as u64,
+                                "retrying after transport error"
+                            );
+                        }
+                        std::thread::sleep(delay);
+                        attempt += 1;
+                        continue;
+                    } else {
+                        return Err(OpenAiError::Http(err));
+                    }
+                }
+            }
         }
     }
 
@@ -68,19 +299,7 @@ impl OpenAiClient {
         );
 
         let url = format!("{}/chat/completions", self.base);
-        let resp = self
-            .http
-            .post(url)
-            .headers(headers)
-            .json(&request)
-            .send()
-            .await?;
-
-        if !resp.status().is_success() {
-            let status = resp.status();
-            let body = resp.text().await.unwrap_or_default();
-            return Err(OpenAiError::Api { status, body });
-        }
+        let resp = self.post_json_with_retry(url, headers, &request).await?;
 
         let bytes = resp.bytes().await?;
         let parsed: ChatCompletionResponse = serde_json::from_slice(&bytes)?;
@@ -110,13 +329,7 @@ impl OpenAiClient {
 
         // 3) async stream wrapper
         try_stream! {
-            let resp = self.http.post(url).headers(headers).json(&request).send().await?;
-
-            if !resp.status().is_success() {
-                let status = resp.status();
-                let body = resp.text().await.unwrap_or_default();
-                return  Err(OpenAiError::Api { status, body })?;
-            }
+            let resp = self.post_json_with_retry(url, headers, &request).await?;
 
             let mut bytes_stream = resp.bytes_stream();
             let mut buf = Vec::new();
diff --git a/crates/artificial-openai/src/error.rs b/crates/artificial-openai/src/error.rs
index 8b19766..70b30dc 100644
--- a/crates/artificial-openai/src/error.rs
+++ b/crates/artificial-openai/src/error.rs
@@ -2,6 +2,18 @@ use std::str::Utf8Error;
 
 use artificial_core::error::ArtificialError;
 use reqwest::StatusCode;
+use std::time::Duration;
+
+/// Headers conveying rate limit information returned by OpenAI.
+#[derive(Debug, Clone)]
+pub struct OpenAiRateLimitHeaders {
+    pub limit_requests: Option<u32>,
+    pub remaining_requests: Option<u32>,
+    pub reset_requests: Option<String>,
+    pub limit_tokens: Option<u32>,
+    pub remaining_tokens: Option<u32>,
+    pub reset_tokens: Option<String>,
+}
 
 /// High-level error type covering every failure mode the client can hit.
 #[derive(Debug, thiserror::Error)]
@@ -12,6 +24,15 @@ pub enum OpenAiError {
     #[error("couldn’t serialise body: {0}")]
     Serde(#[from] serde_json::Error),
 
+    #[error("rate limited (status {status}), retry_after={retry_after:?}, reset_at={reset_at:?}")]
+    RateLimited {
+        status: StatusCode,
+        body: String,
+        retry_after: Option<Duration>,
+        reset_at: Option<String>,
+        headers: OpenAiRateLimitHeaders,
+    },
+
     #[error("OpenAI returned non-success status {status}: {body}")]
     Api { status: StatusCode, body: String },
 
diff --git a/crates/artificial-openai/src/lib.rs b/crates/artificial-openai/src/lib.rs
index 93bcc1e..b43782d 100644
--- a/crates/artificial-openai/src/lib.rs
+++ b/crates/artificial-openai/src/lib.rs
@@ -7,4 +7,5 @@ mod provider_impl_prompt;
 pub use adapter::{OpenAiAdapter, OpenAiAdapterBuilder};
 mod api_v1;
 mod client;
+pub use client::RetryPolicy;
 pub mod error;
diff --git a/crates/artificial-openai/src/provider_impl_chat_stream.rs b/crates/artificial-openai/src/provider_impl_chat_stream.rs
index 3f2d974..a854b08 100644
--- a/crates/artificial-openai/src/provider_impl_chat_stream.rs
+++ b/crates/artificial-openai/src/provider_impl_chat_stream.rs
@@ -1,6 +1,7 @@
 use std::pin::Pin;
 
 use crate::OpenAiAdapter;
+use crate::api_v1::ChatCompletionMessage;
 use crate::api_v1::ChatCompletionRequest;
 use crate::api_v1::FinishReason;
 use artificial_core::error::{ArtificialError, Result};
@@ -11,6 +12,8 @@ use futures_core::stream::Stream;
 use std::collections::HashMap;
 
 impl StreamingChatProvider for OpenAiAdapter {
+    type Message = ChatCompletionMessage;
+
     type Delta<'s>
         = Pin<Box<dyn Stream<Item = Result<String>> + Send + 's>>
     where
diff --git a/crates/artificial-prompt/Cargo.toml b/crates/artificial-prompt/Cargo.toml
index ddaaa7b..3953f15 100644
--- a/crates/artificial-prompt/Cargo.toml
+++ b/crates/artificial-prompt/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "artificial-prompt"
-version = "0.4.2"
+version = "0.5.0"
 edition = "2024"
 description = "Fluent builders and helpers for composing markdown prompt fragments"
 license = "MIT"
@@ -9,4 +9,4 @@ categories = ["development-tools", "text-processing"]
 keywords = ["ai", "prompt", "markdown", "builder"]
 
 [dependencies]
-artificial-core = { path = "../artificial-core" , version = "0.4.2"}
+artificial-core = { path = "../artificial-core" , version = "0.5.0"}
diff --git a/crates/artificial-types/Cargo.toml b/crates/artificial-types/Cargo.toml
index bc69f14..956f1e1 100644
--- a/crates/artificial-types/Cargo.toml
+++ b/crates/artificial-types/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "artificial-types"
-version = "0.4.2"
+version = "0.5.0"
 edition = "2024"
 description = "Reusable prompt fragments and helper types for the Artificial prompt-engineering SDK"
 license = "MIT"
@@ -9,8 +9,8 @@ categories = ["development-tools", "text-processing"]
 keywords = ["ai", "prompt-fragments", "json-schema", "sdk"]
 
 [dependencies]
-artificial-core = { path = "../artificial-core" , version = "0.4.2"}
-artificial-prompt = { path = "../artificial-prompt" , version = "0.4.2"}
+artificial-core = { path = "../artificial-core" , version = "0.5.0"}
+artificial-prompt = { path = "../artificial-prompt" , version = "0.5.0"}
 chrono = "0.4.41"
 
 schemars.workspace = true
diff --git a/crates/artificial/Cargo.toml b/crates/artificial/Cargo.toml
index 7f3cd2a..80a8f84 100644
--- a/crates/artificial/Cargo.toml
+++ b/crates/artificial/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "artificial"
-version = "0.4.2"
+version = "0.5.0"
 edition = "2024"
 description = "Typed, provider-agnostic prompt-engineering SDK for Rust"
 authors = ["Marc Riegel <mail@mrcrgl.de>"]
@@ -13,12 +13,13 @@ keywords = ["ai", "openai", "prompt-engineering", "json-schema", "typed"]
 [features]
 default = ["openai"]
 openai = ["dep:artificial-openai"]
+tracing = ["artificial-openai/tracing"]
 
 [dependencies]
-artificial-types = { path = "../artificial-types", version = "0.4.2" }
-artificial-openai = { path = "../artificial-openai", optional = true, version = "0.4.2" }
-artificial-core = { path = "../artificial-core", version = "0.4.2" }
-artificial-prompt = { path = "../artificial-prompt", version = "0.4.2" }
+artificial-types = { path = "../artificial-types", version = "0.5.0" }
+artificial-openai = { path = "../artificial-openai", optional = true, version = "0.5.0" }
+artificial-core = { path = "../artificial-core", version = "0.5.0" }
+artificial-prompt = { path = "../artificial-prompt", version = "0.5.0" }
 
 [dev-dependencies]
 tokio = { version = "1", features = ["full"] }