feat: extraction models

Cargo.lock

@@ -1496,9 +1496,9 @@ checksum = "8f42a60cbdf9a97f5d2305f08a87dc4e09308d1276d28c869c684d7777685682"
 
 [[package]]
 name = "jiff"
-version = "0.2.37"
+version = "0.2.38"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0ab1baf72f08796de0260609515130699b890ac25f30e610ad894bc5856cafdb"
+checksum = "b2b005715dcbeb0089a3c0dab99f2ff1cc3b2525323552703d648585d342a383"
 dependencies = [
  "defmt",
  "jiff-core",
@@ -1523,9 +1523,9 @@ dependencies = [
 
 [[package]]
 name = "jiff-static"
-version = "0.2.37"
+version = "0.2.38"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "378268a1116ad67ae6228701118ac9f491d78fda38a40a1f1a9e1348de6f7212"
+checksum = "2cc9817253cf7c7ee4684451bd327e88d6f3658014e54a29198625590650695c"
 dependencies = [
  "jiff-core",
  "proc-macro2",
@@ -1633,6 +1633,7 @@ dependencies = [
  "headers",
  "indexmap 2.14.2",
  "itertools 0.15.0",
+ "jiff",
  "mitsein",
  "rand 0.10.3",
  "reqwest",

Cargo.toml

@@ -16,6 +16,7 @@ futures = "0.3.34"
 headers = "0.4.2"
 indexmap = "2.14.2"
 itertools = "0.15.0"
+jiff = { version = "0.2.38", features = ["serde"] }
 mitsein = { version = "0.9.0", features = ["indexmap"] }
 rand = "0.10.3"
 reqwest = { version = "0.13.5", features = ["brotli", "cookies", "deflate", "form", "gzip", "hickory-dns", "json", "multipart", "query", "rustls-no-provider", "socks", "stream", "zstd"], default-features = false }

src/lib.rs

@@ -39,6 +39,9 @@ pub struct AppArgs {
     #[arg(long)]
     pub cloudflare_workers_ai_credentials: Vec<CloudflareCredentials>,
 
+    #[arg(long)]
+    pub cohere_api_key: Vec<SecretString>,
+
     #[arg(long)]
     pub google_ai_studio_api_key: Vec<SecretString>,
 
@@ -72,6 +75,9 @@ pub struct AppArgs {
     #[arg(long)]
     pub openrouter_api_key: Vec<SecretString>,
 
+    #[arg(long)]
+    pub poolside_api_key: Vec<SecretString>,
+
     #[arg(long)]
     pub vercel_ai_gateway_api_key: Vec<SecretString>,
 

src/openai/extraction.rs

@@ -0,0 +1,524 @@
+//! System One-style models that can extract, compute, estimate,
+//! and forecast values from inputs.
+//! They are strongly typed and output confidence scores and probabilities against their outputs.
+//!
+//! Request: [`ExtractionRequest`] · Response: [`ExtractionResponse`]
+
+use jiff::civil::{Date, DateTime, Time};
+use serde::{Deserialize, Serialize};
+use std::collections::HashMap;
+
+/// The item type required for `list` fields.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
+#[serde(rename_all = "lowercase")]
+pub enum ListItems {
+    /// List of text items.
+    Text,
+    /// List of numbers.
+    Number,
+    /// List of ISO dates.
+    Date,
+}
+
+/// Scale for estimate/forecast grids.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq, Default)]
+#[serde(rename_all = "lowercase")]
+pub enum Scale {
+    /// Linear scale (default).
+    #[default]
+    Linear,
+    /// Logarithmic scale.
+    Log,
+}
+
+/// Interval for forecast fields.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
+#[serde(rename_all = "lowercase")]
+pub enum ForecastInterval {
+    /// One step of the grid.
+    Step,
+    /// Per minute.
+    Minute,
+    /// Per hour.
+    Hour,
+    /// Per day.
+    Day,
+    /// Per week.
+    Week,
+    /// Per month.
+    Month,
+    /// Per quarter.
+    Quarter,
+    /// Per year.
+    Year,
+}
+
+/// A value that may be a number or an ISO date string. Used for the
+/// `min`/`max` bounds of estimate/forecast fields (ISO date for date
+/// estimates) and for the `step` grid.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+#[serde(untagged)]
+pub enum NumberOrDate {
+    /// A numeric bound or step size.
+    Number(f64),
+    /// An ISO date string (for date estimates).
+    Date(Date),
+}
+
+/// A step size for estimate grids: a positive number (days for date
+/// estimates), or the strings `"month"` / `"year"`.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+#[serde(untagged)]
+pub enum EstimateStep {
+    /// A positive numeric step (days for date estimates).
+    Number(f64),
+    /// Calendar-aware step: `"month"`.
+    Month,
+    /// Calendar-aware step: `"year"`.
+    Year,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "lowercase")]
+pub enum DurationUnit {
+    Seconds,
+    Minutes,
+    Hours,
+    Days,
+    Weeks,
+    Months,
+    Years,
+}
+
+/// A typed field specification
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+#[serde(tag = "type", rename_all = "lowercase")]
+pub enum Field {
+    /// A string extracted from the input.
+    #[serde(rename = "text")]
+    Text {
+        /// What the field should contain.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        instructions: Option<String>,
+        /// Probability of null above which the value is null (default 0.5).
+        /// Must be in `(0, 1]`.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        null_threshold: Option<f64>,
+    },
+    /// A numeric value.
+    #[serde(rename = "number")]
+    Number {
+        /// What the field should contain.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        instructions: Option<String>,
+        /// Unit label for the number (e.g. `USD`).
+        #[serde(skip_serializing_if = "Option::is_none")]
+        unit: Option<String>,
+        /// Whether the number must be an integer. Numbers and durations only.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        integer: Option<bool>,
+        /// Number of decimal places (0–6). Numbers, durations and number
+        /// lists only.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        decimals: Option<u8>,
+        /// Probability of null above which the value is null (default 0.5).
+        #[serde(skip_serializing_if = "Option::is_none")]
+        null_threshold: Option<f64>,
+    },
+    /// An ISO date value.
+    #[serde(rename = "date")]
+    Date {
+        /// What the field should contain.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        instructions: Option<String>,
+        /// Probability of null above which the value is null (default 0.5).
+        #[serde(skip_serializing_if = "Option::is_none")]
+        null_threshold: Option<f64>,
+    },
+    /// An ISO time value.
+    #[serde(rename = "time")]
+    Time {
+        /// What the field should contain.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        instructions: Option<String>,
+        /// Probability of null above which the value is null (default 0.5).
+        #[serde(skip_serializing_if = "Option::is_none")]
+        null_threshold: Option<f64>,
+    },
+    /// An ISO date-time value.
+    #[serde(rename = "datetime")]
+    Datetime {
+        /// What the field should contain.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        instructions: Option<String>,
+        /// Probability of null above which the value is null (default 0.5).
+        #[serde(skip_serializing_if = "Option::is_none")]
+        null_threshold: Option<f64>,
+    },
+    /// A duration value.
+    #[serde(rename = "duration")]
+    Duration {
+        /// What the field should contain.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        instructions: Option<String>,
+        /// Unit label
+        #[serde(skip_serializing_if = "Option::is_none")]
+        unit: Option<DurationUnit>,
+        /// Whether the duration must be an integer. Numbers and durations only.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        integer: Option<bool>,
+        /// Number of decimal places (0–6). Numbers, durations and number
+        /// lists only.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        decimals: Option<u8>,
+        /// Probability of null above which the value is null (default 0.5).
+        #[serde(skip_serializing_if = "Option::is_none")]
+        null_threshold: Option<f64>,
+    },
+    /// A list of typed items.
+    #[serde(rename = "list")]
+    List {
+        /// What the field should contain.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        instructions: Option<String>,
+        /// Required for list fields: the type of each item.
+        items: ListItems,
+        /// Number of decimal places (0–6) for number lists.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        decimals: Option<u8>,
+        /// Probability of null above which the value is null (default 0.5).
+        #[serde(skip_serializing_if = "Option::is_none")]
+        null_threshold: Option<f64>,
+    },
+    /// A probabilistic estimate over a bounded grid.
+    #[serde(rename = "estimate")]
+    Estimate {
+        /// What the field should contain.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        instructions: Option<String>,
+        /// Required for estimate/forecast: lower bound (a number, or an ISO
+        /// date for date estimates).
+        min: NumberOrDate,
+        /// Required for estimate/forecast: upper bound (a number, or an ISO
+        /// date for date estimates).
+        max: NumberOrDate,
+        /// Estimate grid: a positive number (days for date estimates), or
+        /// `"month"` / `"year"`.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        step: Option<EstimateStep>,
+        /// Grid scale: `linear` (default) or `log`.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        scale: Option<Scale>,
+        /// Probability of null above which the value is null (default 0.5).
+        #[serde(skip_serializing_if = "Option::is_none")]
+        null_threshold: Option<f64>,
+    },
+    /// A forecast: a sequence of estimates over a horizon.
+    #[serde(rename = "forecast")]
+    Forecast {
+        /// What the field should contain.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        instructions: Option<String>,
+        /// Required for estimate/forecast: lower bound (a number, or an ISO
+        /// date for date estimates).
+        min: NumberOrDate,
+        /// Required for estimate/forecast: upper bound (a number, or an ISO
+        /// date for date estimates).
+        max: NumberOrDate,
+        /// Estimate grid: a positive number (days for date estimates), or
+        /// `"month"` / `"year"`.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        step: Option<EstimateStep>,
+        /// Grid scale: `linear` (default) or `log`.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        scale: Option<Scale>,
+        /// Forecast only: number of steps (1–64).
+        #[serde(skip_serializing_if = "Option::is_none")]
+        horizon: Option<u32>,
+        /// Forecast only: the interval between steps.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        interval: Option<ForecastInterval>,
+        /// Probability of null above which the value is null (default 0.5).
+        #[serde(skip_serializing_if = "Option::is_none")]
+        null_threshold: Option<f64>,
+    },
+}
+
+/// An image input. Either a data URL string or an explicit base64 object.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+#[serde(untagged)]
+pub enum Image {
+    /// A data URL string (must start with `data:`, case-insensitive).
+    DataUrl(
+        /// The data URL, e.g. `data:image/png;base64,...`.
+        String,
+    ),
+    /// An explicitly split base64 image.
+    Base64 {
+        /// MIME content type of the image (PNG, JPEG or WebP).
+        content_type: String,
+        /// Base64-encoded image bytes.
+        base64: String,
+    },
+}
+
+/// Optional request settings.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Default)]
+pub struct RequestOptions {
+    /// Add 99 percentiles to estimates (default `false`).
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub include_quantiles: Option<bool>,
+    /// Return the text that spans index into (default `false`).
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub return_state_text: Option<bool>,
+}
+
+/// The extraction request body.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub struct ExtractionRequest<Model> {
+    /// The model selector
+    pub model: Model,
+    /// The text to read and predict from. Must be non-empty unless images
+    /// are sent. Text answers are spans of this string.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub state: Option<String>,
+    /// Required. Map of field name to a typed field spec. At least one
+    /// field; no upper limit. Fields that do not fit next to the state in
+    /// one 8,192-token pass are split across parallel passes. Field names
+    /// must match `^[A-Za-z_][A-Za-z0-9_.\-]{0,63}$`.
+    pub fields: HashMap<String, Field>,
+    /// Optional embedded PNG, JPEG or WebP images placed before the state
+    /// (max 4; 4 MiB each, 8 MiB total decoded; request body max 16 MiB).
+    /// Remote URLs are not accepted.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub images: Option<Vec<Image>>,
+    /// ISO date that anchors dates written without a year.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub reference_date: Option<Date>,
+    /// Optional request settings.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub options: Option<RequestOptions>,
+}
+
+/// The value of an answer. Shape depends on the field type; `null` when the
+/// input does not determine it.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+#[serde(untagged)]
+pub enum List {
+    // A list of strings
+    Strings(Vec<String>),
+    // A list of numbers
+    Numbers(Vec<f64>),
+    // A list of dates
+    Dates(Vec<Date>),
+}
+
+/// A forecast step: one estimate per step, with median and interval bounds.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub struct ForecastStep {
+    /// The point (median) value for this step.
+    pub value: f64,
+    /// Mean of the distribution for this step.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub mean: Option<f64>,
+    /// Lower bound of the interval for this step.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub p10: Option<f64>,
+    /// Upper bound of the interval for this step.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub p90: Option<f64>,
+}
+
+/// One answer per requested field, keyed by field name. Tagged by `type`,
+/// mirroring the field spec that produced it.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+#[serde(tag = "type", rename_all = "lowercase")]
+pub enum Answer {
+    /// Answer for a text field: an exact span of the state, with character
+    /// offsets.
+    #[serde(rename = "text")]
+    Text {
+        /// The extracted value, or null when the input does not determine it.
+        value: Option<String>,
+        /// The model's best guess before applying the null threshold.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        guess: Option<String>,
+        /// Probability that the field is null.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        p_null: Option<f64>,
+        /// Confidence score in `[0, 1]`.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        confidence: Option<f64>,
+        /// Character offsets `[start, end]` of the span into the state.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        span: Option<[usize; 2]>,
+    },
+    /// Answer for a number field.
+    #[serde(rename = "number")]
+    Number {
+        /// The extracted value, or null when the input does not determine it.
+        value: Option<f64>,
+        /// The model's best guess before applying the null threshold.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        guess: Option<f64>,
+        /// Probability that the field is null.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        p_null: Option<f64>,
+        /// Confidence score in `[0, 1]`.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        confidence: Option<f64>,
+    },
+    /// Answer for a date field.
+    #[serde(rename = "date")]
+    Date {
+        /// The computed/extracted ISO date, or null.
+        value: Option<Date>,
+        /// The model's best guess before applying the null threshold.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        guess: Option<Date>,
+        /// Probability that the field is null.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        p_null: Option<f64>,
+        /// Confidence score in `[0, 1]`.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        confidence: Option<f64>,
+    },
+    /// Answer for a time field.
+    #[serde(rename = "time")]
+    Time {
+        /// The extracted ISO time, or null.
+        value: Option<Time>,
+        /// The model's best guess before applying the null threshold.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        guess: Option<Time>,
+        /// Probability that the field is null.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        p_null: Option<f64>,
+        /// Confidence score in `[0, 1]`.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        confidence: Option<f64>,
+    },
+    /// Answer for a datetime field.
+    #[serde(rename = "datetime")]
+    Datetime {
+        /// The extracted ISO date-time, or null.
+        value: Option<DateTime>,
+        /// The model's best guess before applying the null threshold.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        guess: Option<DateTime>,
+        /// Probability that the field is null.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        p_null: Option<f64>,
+        /// Confidence score in `[0, 1]`.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        confidence: Option<f64>,
+    },
+    /// Answer for a duration field.
+    #[serde(rename = "duration")]
+    Duration {
+        /// The extracted duration, or null.
+        value: Option<f64>,
+        /// The model's best guess before applying the null threshold.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        guess: Option<f64>,
+        /// Probability that the field is null.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        p_null: Option<f64>,
+        /// Confidence score in `[0, 1]`.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        confidence: Option<f64>,
+    },
+    /// Answer for a list field: values with per-item spans.
+    #[serde(rename = "list")]
+    List {
+        /// The extracted items, or null.
+        value: Option<List>,
+        /// Per-item character offsets `[start, end]` into the state.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        spans: Option<Vec<[usize; 2]>>,
+        /// Probability that the field is null.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        p_null: Option<f64>,
+        /// Confidence score in `[0, 1]`.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        confidence: Option<f64>,
+    },
+    /// Answer for an estimate field: a distribution summary.
+    #[serde(rename = "estimate")]
+    Estimate {
+        /// The point estimate (median).
+        value: NumberOrDate,
+        /// Mean of the estimated distribution.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        mean: Option<f64>,
+        /// Median of the estimated distribution.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        median: Option<f64>,
+        /// 10th percentile of the estimated distribution.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        p10: Option<f64>,
+        /// 90th percentile of the estimated distribution.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        p90: Option<f64>,
+        /// Probability the true value is below the grid minimum.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        p_below: Option<f64>,
+        /// Probability the true value is above the grid maximum.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        p_above: Option<f64>,
+        /// 99 percentiles, when `options.include_quantiles` is set.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        quantiles: Option<Vec<f64>>,
+    },
+    /// Answer for a forecast field: one estimate per step.
+    #[serde(rename = "forecast")]
+    Forecast {
+        /// One estimate per step.
+        #[serde(default)]
+        steps: Vec<ForecastStep>,
+        /// Probability that the field is null.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        p_null: Option<f64>,
+        /// Confidence score in `[0, 1]`.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        confidence: Option<f64>,
+    },
+}
+
+/// Token and pass accounting for a request.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub struct Usage {
+    /// Tokens consumed by the state text.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub state_tokens: Option<u64>,
+    /// Tokens consumed by images.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub image_tokens: Option<u64>,
+    /// Tokens consumed by the field schema.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub schema_tokens: Option<u64>,
+    /// Total input tokens.
+    pub input_tokens: u64,
+    /// Total output tokens.
+    pub output_tokens: u64,
+    /// Number of parallel passes used.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub passes: Option<u64>,
+}
+
+/// The extraction response body.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub struct ExtractionResponse<Model> {
+    /// The model that produced this response
+    pub model: Model,
+    /// One answer per field, keyed by field name. `value` is null when the
+    /// input does not determine it.
+    pub answers: HashMap<String, Answer>,
+    /// Whether the input was truncated to fit the context window.
+    pub truncated: bool,
+    /// The text that spans index into, when `options.return_state_text` is set.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub state_text: Option<String>,
+    /// Token and pass accounting for the request.
+    pub usage: Usage,
+}

src/openai/mod.rs

@@ -1,5 +1,6 @@
 pub mod chat;
 pub mod decisions;
 pub mod embeddings;
+pub mod extraction;
 pub mod models;
 pub mod system_one;

src/server/chat/completions.rs

@@ -25,9 +25,12 @@ use std::collections::VecDeque;
 use std::convert::Infallible;
 use std::fmt::Display;
 use std::str::FromStr;
+use std::time::Duration;
 use strum::EnumString;
 use strum::VariantArray;
 use tokio::runtime::Handle;
+use tokio::select;
+use tokio::time::sleep;
 
 #[derive(Deserialize, TypedPath)]
 #[typed_path("/v1/chat/completions")]
@@ -88,6 +91,7 @@ impl Display for ShortcutAndMaybeReasoning {
 #[derive(Debug, Clone)]
 enum Provider {
     CloudflareWorkersAi,
+    Cohere,
     GoogleAiStudio,
     Groq,
     HuggingFace,
@@ -98,6 +102,7 @@ enum Provider {
     Ollama,
     OpenCode,
     OpenRouter,
+    Poolside,
     VercelAiGateway,
     ZAi,
 }
@@ -110,19 +115,22 @@ impl Shortcut {
         match self {
             S::XS => vec![
                 (P::GoogleAiStudio, "gemini-3.5-flash-lite".into()),
-                (P::NovitaAi, "apodex/apodex-1.1-mini".into()),
-                (P::OpenRouter, "apodex/apodex-1.1-mini:free".into()),
                 (P::VercelAiGateway, "openai/gpt-5.4-nano".into()),
                 (P::HuggingFace, "Qwen/Qwen3.6-35B-A3B".into()),
+                (P::OpenRouter, "xingchen-agi/xing4.0-29b-a4b:free".into()),
+                (P::NovitaAi, "apodex/apodex-1.1-mini".into()),
+                (P::OpenRouter, "apodex/apodex-1.1-mini:free".into()),
                 (P::VercelAiGateway, "liquid/lfm2.5-24b-a2b".into()),
                 (P::GoogleAiStudio, "gemma-4-26b-a4b-it".into()),
                 (P::OpenRouter, "google/gemma-4-26b-a4b-it:free".into()),
                 (P::Kilo, "google-gemma-4-26b-a4b-it-free".into()),
                 (P::HuggingFace, "google/gemma-4-26B-A4B-it".into()),
+                (P::Poolside, "poolside/laguna-xs-2.1".into()),
                 (P::NvidiaNim, "poolside/laguna-xs-2.1".into()),
                 (P::OpenRouter, "poolside/laguna-xs-2.1:free".into()),
                 (P::ZAi, "glm-4.7-flash".into()),
                 (P::NvidiaNim, "google/diffusiongemma-26b-a4b-it".into()),
+                (P::Cohere, "north-mini-code-1-0".into()),
                 (P::OpenRouter, "cohere/north-mini-code:free".into()),
                 (P::NvidiaNim, "nvidia/nemotron-3.5-lightning-30b-a3b".into()),
                 (
@@ -185,28 +193,21 @@ impl Shortcut {
                 (P::OpenCode, "nemotron-3.5-super-vl-free".into()),
                 (P::Mistral, "mistral-medium-3.5".into()),
                 (P::Kilo, "inclusionai-ling-3-0-flash-free".into()),
-                (P::NovitaAi, "inclusionai/ling-3.0-flash-vl".into()),
                 (P::Kilo, "inclusionai-ling-3-0-flash-vl-free".into()),
-                (P::OpenRouter, "inclusionai/ling-3.0-flash-vl:free".into()),
                 (P::VercelAiGateway, "inclusionai/ling-3.0-flash-vl".into()),
                 (P::VercelAiGateway, "poolside/laguna-s-2.1-free".into()),
+                (P::Poolside, "poolside/laguna-s-2.1".into()),
                 (P::Kilo, "poolside-laguna-s-2-1-free".into()),
                 (P::OpenRouter, "poolside/laguna-s-2.1:free".into()),
                 (P::VercelAiGateway, "stepfun/step-3.7-flash".into()),
-                (P::NovitaAi, "inclusionai/ling-3.0-flash-fin".into()),
-                (P::OpenRouter, "inclusionai/ling-3.0-flash-fin:free".into()),
                 (P::VercelAiGateway, "inclusionai/ling-3.0-flash-fin".into()),
-                (P::NovitaAi, "inclusionai/ling-3.0-flash-sante".into()),
-                (
-                    P::OpenRouter,
-                    "inclusionai/ling-3.0-flash-sante:free".into(),
-                ),
                 (
                     P::VercelAiGateway,
                     "inclusionai/ling-3.0-flash-sante".into(),
                 ),
                 (P::VercelAiGateway, "mistral/mistral-large-3".into()),
                 (P::Ollama, "nemotron-3-super".into()),
+                (P::Cohere, "command-a-plus-05-2026".into()),
                 (P::NvidiaNim, "nvidia/nemotron-3-super-120b-a12b".into()),
                 (P::Kilo, "nvidia-nemotron-3-super-120b-a12b-free".into()),
                 (
@@ -233,8 +234,11 @@ impl Shortcut {
                 (P::NvidiaNim, "deepseek-ai/deepseek-v4.1-flash".into()),
                 (P::HuggingFace, "deepseek-ai/DeepSeek-V4.1-Flash".into()),
                 (P::OpenCode, "deepseek-v4.1-flash-free".into()),
+                (P::VercelAiGateway, "minimax/minimax-m3.1-flash".into()),
+                (P::OpenRouter, "minimax/minimax-m3.1-flash:free".into()),
                 (P::VercelAiGateway, "inclusionai/ling-3.1-flash".into()),
                 (P::HuggingFace, "inclusionAI/Ling-3.1-flash".into()),
+                (P::VercelAiGateway, "reflection/beam-free".into()),
                 (P::Ollama, "beam".into()),
                 (P::OpenCode, "beam-free".into()),
                 (P::VercelAiGateway, "xiaomi/mimo-v2.6-flash".into()),
@@ -254,6 +258,10 @@ impl Shortcut {
                 (P::Ollama, "nemotron-3-ultra".into()),
                 (P::NvidiaNim, "nvidia/nemotron-3-ultra-550b-a55b".into()),
                 (P::Kilo, "nvidia-nemotron-3-ultra-550b-a55b-free".into()),
+                (
+                    P::OpenRouter,
+                    "nvidia/nemotron-3-ultra-550b-a55b:free".into(),
+                ),
                 (P::OpenCode, "nemotron-3-ultra-free".into()),
             ],
 
@@ -262,6 +270,14 @@ impl Shortcut {
                 (P::HuggingFace, "zai-org/GLM-5.3".into()),
                 (P::CloudflareWorkersAi, "@cf/zai-org/glm-5.3".into()),
                 (P::Mistral, "zai-glm-5-3".into()),
+                (
+                    P::VercelAiGateway,
+                    "deepseek/deepseek-v4.1-flash-1017".into(),
+                ),
+                (
+                    P::CloudflareWorkersAi,
+                    "@cf/deepseek-ai/deepseek-v4.1-flash-1017".into(),
+                ),
                 (P::OpenCode, "muse-spark-1.3-contributor-free".into()),
                 (P::NvidiaNim, "moonshotai/kimi-k3".into()),
                 (P::HuggingFace, "moonshotai/Kimi-K3".into()),
@@ -395,6 +411,22 @@ pub async fn create_chat_completion(
                             .build()
                     })
                     .collect(),
+                Provider::Cohere => state
+                    .app_args
+                    .cohere_api_key
+                    .iter()
+                    .cloned()
+                    .map(|api_key| {
+                        Upstream::builder()
+                            .url(
+                                "https://api.cohere.com/compatibility/v1/chat/completions"
+                                    .parse()
+                                    .expect("TODO"),
+                            )
+                            .api_key(api_key)
+                            .build()
+                    })
+                    .collect(),
                 Provider::Groq => state
                     .app_args
                     .groq_api_key
@@ -503,6 +535,18 @@ pub async fn create_chat_completion(
                             .build()
                     })
                     .collect(),
+                Provider::Poolside => state
+                    .app_args
+                    .poolside_api_key
+                    .iter()
+                    .cloned()
+                    .map(|api_key| {
+                        Upstream::builder()
+                            .url("https://inference.poolside.ai/v1/chat/completions".parse().expect("TODO"))
+                            .api_key(api_key)
+                            .build()
+                    })
+                    .collect(),
                 Provider::VercelAiGateway => state
                     .app_args
                     .vercel_ai_gateway_api_key
@@ -649,12 +693,21 @@ pub async fn create_chat_completion(
             });
         }
 
-        let upstream_response = match upstream_request.send().await {
+        let upstream_response = select! {
-            Ok(upstream_response) => upstream_response,
+            _ = sleep(Duration::from_secs(20)) => {
-            Err(upstream_error) => {
+                tracing::error!("timed out");
-                tracing::error!(?upstream_error);
                 continue;
             }
+
+            upstream_response_result = upstream_request.send() => {
+                match upstream_response_result {
+                    Ok(upstream_response) => upstream_response,
+                    Err(upstream_error) => {
+                        tracing::error!(?upstream_error);
+                        continue;
+                    }
+                }
+            }
         };
 
         if let Err(upstream_error) = upstream_response.error_for_status_ref() {

src/server/embeddings.rs

@@ -1,4 +1,3 @@
-use crate::CloudflareCredentials;
 use crate::openai;
 use crate::server::AppState;
 use crate::server::Upstream;
@@ -34,6 +33,18 @@ pub enum Shortcut {
     #[strum(serialize = "codestral-embed-2505")]
     CodestralEmbed2505,
 
+    #[strum(serialize = "cohere-embed-v4.0")]
+    CohereEmbedV4_0,
+
+    #[strum(serialize = "cohere-embed-v5.0-fast")]
+    CohereEmbedV5_0Fast,
+
+    #[strum(serialize = "cohere-embed-v5.0-pro")]
+    CohereEmbedV5_0Pro,
+
+    #[strum(serialize = "gemini-embedding-2")]
+    GeminiEmbedding2,
+
     #[strum(serialize = "lfm2.5-embedding-350m")]
     Lfm2_5Embedding350M,
 
@@ -46,9 +57,12 @@ pub enum Shortcut {
 
 #[derive(Debug, Clone)]
 enum Provider {
+    Cohere,
+    GoogleAiStudio,
     Mistral,
     NvidiaNim,
     OpenRouter,
+    VercelAiGateway,
 }
 
 impl Shortcut {
@@ -58,6 +72,19 @@ impl Shortcut {
 
         match self {
             S::CodestralEmbed2505 => vec![(P::Mistral, "codestral-embed-2505".into())],
+            S::CohereEmbedV4_0 => vec![
+                (P::Cohere, "embed-v4.0".into()),
+                (P::VercelAiGateway, "cohere/embed-v4.0".into()),
+            ],
+            S::CohereEmbedV5_0Fast => vec![
+                (P::Cohere, "embed-v5.0-fast".into()),
+                (P::VercelAiGateway, "cohere/embed-v5.0-fast".into()),
+            ],
+            S::CohereEmbedV5_0Pro => vec![
+                (P::Cohere, "embed-v5.0-pro".into()),
+                (P::VercelAiGateway, "cohere/embed-v5.0-pro".into()),
+            ],
+            S::GeminiEmbedding2 => vec![(P::GoogleAiStudio, "gemini-embedding-2".into())],
             S::Lfm2_5Embedding350M => {
                 vec![(P::OpenRouter, "liquid/lfm-2.5-embedding-350m:free".into())]
             }
@@ -103,6 +130,22 @@ pub async fn create_embeddings(
 
     for (provider, model) in shortcut.suggestions() {
         let upstreams: Vec<Upstream> = match provider {
+            Provider::Cohere => state
+                .app_args
+                .cohere_api_key
+                .iter()
+                .cloned()
+                .map(|api_key| {
+                    Upstream::builder()
+                        .url(
+                            "https://api.cohere.com/compatibility/v1/embeddings"
+                                .parse()
+                                .expect("TODO"),
+                        )
+                        .api_key(api_key)
+                        .build()
+                })
+                .collect(),
             Provider::Mistral => state
                 .app_args
                 .mistral_api_key
@@ -119,6 +162,22 @@ pub async fn create_embeddings(
                         .build()
                 })
                 .collect(),
+            Provider::GoogleAiStudio => state
+                .app_args
+                .google_ai_studio_api_key
+                .iter()
+                .cloned()
+                .map(|api_key| {
+                    Upstream::builder()
+                        .url(
+                            "https://generativelanguage.googleapis.com/v1beta/openai/embeddings"
+                                .parse()
+                                .expect("TODO"),
+                        )
+                        .api_key(api_key)
+                        .build()
+                })
+                .collect(),
             Provider::NvidiaNim => state
                 .app_args
                 .nvidia_nim_api_key
@@ -151,6 +210,22 @@ pub async fn create_embeddings(
                         .build()
                 })
                 .collect(),
+            Provider::VercelAiGateway => state
+                .app_args
+                .vercel_ai_gateway_api_key
+                .iter()
+                .cloned()
+                .map(|api_key| {
+                    Upstream::builder()
+                        .url(
+                            "https://ai-gateway.vercel.sh/v1/embeddings"
+                                .parse()
+                                .expect("TODO"),
+                        )
+                        .api_key(api_key)
+                        .build()
+                })
+                .collect(),
         };
 
         if let Ok(upstreams) = Vec1::try_from(upstreams) {

src/server/extraction.rs

@@ -0,0 +1,217 @@
+use crate::CloudflareCredentials;
+use crate::openai;
+use crate::server::AppState;
+use crate::server::Upstream;
+use crate::server::take_random_by_swapping;
+use axum::Json;
+use axum::extract::State;
+use axum::http::StatusCode;
+use axum::response::IntoResponse;
+use axum::response::Response;
+use axum_extra::routing::TypedPath;
+use ecow::EcoString;
+use headers::Header;
+use headers::authorization::Bearer;
+use mitsein::vec1::Vec1;
+use reqwest::Certificate;
+use secrecy::ExposeSecret;
+use serde::Deserialize;
+use serde_with::DeserializeFromStr;
+use serde_with::SerializeDisplay;
+use std::collections::VecDeque;
+use strum::EnumString;
+use strum::VariantArray;
+
+#[derive(Deserialize, TypedPath)]
+#[typed_path("/v1/extraction")]
+pub struct V1ExtractionPath;
+
+#[derive(
+    Clone, Debug, EnumString, DeserializeFromStr, strum::Display, SerializeDisplay, VariantArray,
+)]
+#[strum(serialize_all = "lowercase")]
+pub enum Shortcut {
+    M,
+}
+
+#[derive(Debug, Clone)]
+enum Provider {
+    Cloudflare,
+}
+
+impl Shortcut {
+    fn suggestions(&self) -> Vec<(Provider, EcoString)> {
+        use Provider as P;
+        use Shortcut as S;
+
+        match self {
+            S::M => {
+                vec![(P::Cloudflare, "@cf/cloudflare/clairic".into())]
+            }
+        }
+    }
+}
+
+pub struct ExtractionResponse(pub openai::extraction::ExtractionResponse<String>);
+
+impl IntoResponse for ExtractionResponse {
+    fn into_response(self) -> Response {
+        let Self(extraction) = self;
+
+        Json(extraction).into_response()
+    }
+}
+
+pub enum ExtractionError {
+    NoneSucceeded,
+}
+
+impl IntoResponse for ExtractionError {
+    fn into_response(self) -> Response {
+        match self {
+            ExtractionError::NoneSucceeded => StatusCode::SERVICE_UNAVAILABLE.into_response(),
+        }
+    }
+}
+
+#[tracing::instrument(skip(state, request))]
+pub async fn create_extraction(
+    _: V1ExtractionPath,
+    State(state): State<AppState>,
+    Json(request): Json<openai::extraction::ExtractionRequest<Shortcut>>,
+) -> Result<ExtractionResponse, ExtractionError> {
+    let shortcut = request.model.clone();
+    let mut model_upstreams = VecDeque::new();
+
+    for (provider, model) in shortcut.suggestions() {
+        let upstreams: Vec<Upstream> =
+            match provider {
+                Provider::Cloudflare => state
+                    .app_args
+                    .cloudflare_workers_ai_credentials
+                    .iter()
+                    .cloned()
+                    .map(
+                        |CloudflareCredentials {
+                             account_id,
+                             api_key,
+                         }| {
+                            Upstream::builder()
+                                .url(
+                                    format!("https://api.cloudflare.com/client/v4/accounts/{account_id}/ai/run/{model}")
+                                        .parse()
+                                        .expect("TODO"),
+                                )
+                                .api_key(api_key)
+                                .build()
+                        },
+                    )
+                    .collect(),
+            };
+
+        if let Ok(upstreams) = Vec1::try_from(upstreams) {
+            model_upstreams.push_back((model, upstreams));
+        }
+    }
+
+    while let Some((model, upstreams)) = model_upstreams.pop_front() {
+        let (upstream, upstreams) = take_random_by_swapping(upstreams, &mut rand::rng());
+        if let Some(upstreams) = upstreams {
+            model_upstreams.push_back((model.clone(), upstreams));
+        }
+
+        tracing::info!(?model, ?upstream);
+
+        let openai::extraction::ExtractionRequest {
+            model: _shortcut,
+            state,
+            fields,
+            images,
+            reference_date,
+            options,
+        } = request.clone();
+
+        let request = openai::extraction::ExtractionRequest {
+            model,
+            state,
+            fields,
+            images,
+            reference_date,
+            options,
+        };
+
+        let client = reqwest::Client::builder()
+            .tls_certs_only(
+                webpki_root_certs::TLS_SERVER_ROOT_CERTS
+                    .iter()
+                    .filter_map(|cert| Certificate::from_der(cert).ok()),
+            )
+            .build()
+            .expect("TODO");
+
+        let mut upstream_request = client.post(upstream.url).json(&request);
+
+        if let Some(api_key) = upstream.api_key {
+            let authorization =
+                headers::Authorization::bearer(api_key.expose_secret()).expect("TODO");
+            upstream_request = upstream_request.header(headers::Authorization::<Bearer>::name(), {
+                let mut vec = vec![];
+
+                authorization.encode(&mut vec);
+
+                let [one] = vec.try_into().expect("TODO");
+
+                one
+            });
+        }
+
+        let upstream_response = match upstream_request.send().await {
+            Ok(upstream_response) => upstream_response,
+            Err(upstream_error) => {
+                tracing::error!(?upstream_error);
+                continue;
+            }
+        };
+
+        if let Err(upstream_error) = upstream_response.error_for_status_ref() {
+            tracing::error!(?upstream_error);
+            match upstream_response.text().await {
+                Ok(body) => tracing::error!(?body),
+                Err(body_error) => tracing::error!(?body_error),
+            };
+            continue;
+        };
+
+        let response: serde_json::Value = match upstream_response.json().await {
+            Ok(response) => response,
+            Err(json_error) => {
+                tracing::error!(?json_error);
+                continue;
+            }
+        };
+
+        match openai::extraction::ExtractionResponse::deserialize(
+            response.clone(), // TODO: not need to clone
+        ) {
+            Ok(response) => return Ok(ExtractionResponse(response)),
+            Err(structure_error) => {
+                #[derive(Deserialize)]
+                struct CloudflareRunResponse {
+                    result: openai::extraction::ExtractionResponse<String>,
+                    success: bool,
+                    errors: Vec<serde_json::Value>,   // TODO
+                    messages: Vec<serde_json::Value>, // TODO
+                }
+
+                match CloudflareRunResponse::deserialize(response) {
+                    Ok(response) => return Ok(ExtractionResponse(response.result)),
+                    Err(cloudflare_structure_error) => {
+                        tracing::error!(?structure_error);
+                        tracing::error!(?cloudflare_structure_error);
+                    }
+                }
+            }
+        }
+    }
+    Err(ExtractionError::NoneSucceeded)
+}

src/server/frontend/index.html

@@ -0,0 +1,12 @@
+<!doctype html>
+<html lang="en">
+  <head>
+    <meta charset="utf-8" />
+    <meta name="viewport" content="width=device-width, initial-scale=1" />
+    <title>Basic Page</title>
+  </head>
+  <body>
+    <h1>Hello, world!</h1>
+    <p>This is a basic HTML page.</p>
+  </body>
+</html>

src/server/mod.rs

@@ -12,8 +12,9 @@ use url::Url;
 
 mod chat;
 mod decisions;
-mod frontend;
 mod embeddings;
+mod extraction;
+mod frontend;
 mod models;
 mod system_one;
 
@@ -46,6 +47,7 @@ pub fn create_router() -> Router<AppState> {
         .typed_post(chat::completions::create_chat_completion)
         .typed_post(decisions::create_decision)
         .typed_post(embeddings::create_embeddings)
+        .typed_post(extraction::create_extraction)
         .typed_post(system_one::create_system_one)
         .typed_get(models::list_models)
         .typed_get(frontend::spa)

src/server/system_one.rs

@@ -33,6 +33,7 @@ pub struct V1SystemOnePath;
 pub enum Shortcut {
     S,
     M,
+    L,
 }
 
 #[derive(Debug, Clone)]
@@ -52,6 +53,10 @@ impl Shortcut {
         match self {
             S::S => {
                 vec![
+                    (
+                        P::VercelAiGateway,
+                        "perplexity/pplx-decider-flash-v1.2".into(),
+                    ),
                     (P::LiquidAi, "d1:free".into()),
                     (P::CloudflareWorkersAi, "clef-omni".into()),
                     (P::CloudflareWorkersAi, "clef-flash".into()),
@@ -65,10 +70,15 @@ impl Shortcut {
             S::M => {
                 vec![
                     (P::OpenCode, "jev-1.13-free".into()),
+                    (P::VercelAiGateway, "perplexity/pplx-decider-v1.2".into()),
                     (P::CloudflareWorkersAi, "clef".into()),
                     (P::OpenRouter, "inception/mercury-decide:free".into()),
                 ]
             }
+
+            S::L => {
+                vec![(P::VercelAiGateway, "liquid/d1-ultra".into())]
+            }
         }
     }
 }