Extract completion provider crate (#14823)

We will soon need `semantic_index` to be able to use
`CompletionProvider`. This is currently impossible due to a cyclic crate
dependency, because `CompletionProvider` lives in the `assistant` crate,
which depends on `semantic_index`.

This PR breaks the dependency cycle by extracting two crates out of
`assistant`: `language_model` and `completion`.

Only one piece of logic changed: [this
code](922fcaf5a6 (diff-3857b3707687a4d585f1200eec4c34a7a079eae8d303b4ce5b4fce46234ace9fR61-R69)).
* As of https://github.com/zed-industries/zed/pull/13276, whenever we
ask a given completion provider for its available models, OpenAI
providers would go and ask the global assistant settings whether the
user had configured an `available_models` setting, and if so, return
that.
* This PR changes it so that instead of eagerly asking the assistant
settings for this info (the new crate must not depend on `assistant`, or
else the dependency cycle would be back), OpenAI completion providers
now store the user-configured settings as part of their struct, and
whenever the settings change, we update the provider.

In theory, this change should not change user-visible behavior...but
since it's the only change in this large PR that's more than just moving
code around, I'm mentioning it here in case there's an unexpected
regression in practice! (cc @amtoaer in case you'd like to try out this
branch and verify that the feature is still working the way you expect.)

Release Notes:

- N/A

---------

Co-authored-by: Marshall Bowers <elliott.codes@gmail.com>

crates/language_model/Cargo.toml

@@ -0,0 +1,41 @@
+[package]
+name = "language_model"
+version = "0.1.0"
+edition = "2021"
+publish = false
+license = "GPL-3.0-or-later"
+
+[lints]
+workspace = true
+
+[lib]
+path = "src/language_model.rs"
+doctest = false
+
+[features]
+test-support = [
+    "editor/test-support",
+    "language/test-support",
+    "project/test-support",
+    "text/test-support",
+]
+
+[dependencies]
+anthropic = { workspace = true, features = ["schemars"] }
+ollama = { workspace = true, features = ["schemars"] }
+open_ai = { workspace = true, features = ["schemars"] }
+schemars.workspace = true
+serde.workspace = true
+strum.workspace = true
+proto = { workspace = true, features = ["test-support"] }
+
+[dev-dependencies]
+ctor.workspace = true
+editor = { workspace = true, features = ["test-support"] }
+env_logger.workspace = true
+language = { workspace = true, features = ["test-support"] }
+log.workspace = true
+project = { workspace = true, features = ["test-support"] }
+rand.workspace = true
+text = { workspace = true, features = ["test-support"] }
+unindent.workspace = true

crates/language_model/LICENSE-GPL

@@ -0,0 +1 @@
+../../LICENSE-GPL

crates/language_model/src/language_model.rs

@@ -0,0 +1,7 @@
+mod model;
+mod request;
+mod role;
+
+pub use model::*;
+pub use request::*;
+pub use role::*;

crates/language_model/src/model/cloud_model.rs

@@ -0,0 +1,160 @@
+use crate::LanguageModelRequest;
+pub use anthropic::Model as AnthropicModel;
+pub use ollama::Model as OllamaModel;
+pub use open_ai::Model as OpenAiModel;
+use schemars::{
+    schema::{InstanceType, Metadata, Schema, SchemaObject},
+    JsonSchema,
+};
+use serde::{
+    de::{self, Visitor},
+    Deserialize, Deserializer, Serialize, Serializer,
+};
+use std::fmt;
+use strum::{EnumIter, IntoEnumIterator};
+
+#[derive(Clone, Debug, Default, PartialEq, EnumIter)]
+pub enum CloudModel {
+    Gpt3Point5Turbo,
+    Gpt4,
+    Gpt4Turbo,
+    #[default]
+    Gpt4Omni,
+    Gpt4OmniMini,
+    Claude3_5Sonnet,
+    Claude3Opus,
+    Claude3Sonnet,
+    Claude3Haiku,
+    Gemini15Pro,
+    Gemini15Flash,
+    Custom(String),
+}
+
+impl Serialize for CloudModel {
+    fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
+    where
+        S: Serializer,
+    {
+        serializer.serialize_str(self.id())
+    }
+}
+
+impl<'de> Deserialize<'de> for CloudModel {
+    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
+    where
+        D: Deserializer<'de>,
+    {
+        struct ZedDotDevModelVisitor;
+
+        impl<'de> Visitor<'de> for ZedDotDevModelVisitor {
+            type Value = CloudModel;
+
+            fn expecting(&self, formatter: &mut fmt::Formatter) -> fmt::Result {
+                formatter.write_str("a string for a ZedDotDevModel variant or a custom model")
+            }
+
+            fn visit_str<E>(self, value: &str) -> Result<Self::Value, E>
+            where
+                E: de::Error,
+            {
+                let model = CloudModel::iter()
+                    .find(|model| model.id() == value)
+                    .unwrap_or_else(|| CloudModel::Custom(value.to_string()));
+                Ok(model)
+            }
+        }
+
+        deserializer.deserialize_str(ZedDotDevModelVisitor)
+    }
+}
+
+impl JsonSchema for CloudModel {
+    fn schema_name() -> String {
+        "ZedDotDevModel".to_owned()
+    }
+
+    fn json_schema(_generator: &mut schemars::gen::SchemaGenerator) -> Schema {
+        let variants = CloudModel::iter()
+            .filter_map(|model| {
+                let id = model.id();
+                if id.is_empty() {
+                    None
+                } else {
+                    Some(id.to_string())
+                }
+            })
+            .collect::<Vec<_>>();
+        Schema::Object(SchemaObject {
+            instance_type: Some(InstanceType::String.into()),
+            enum_values: Some(variants.iter().map(|s| s.clone().into()).collect()),
+            metadata: Some(Box::new(Metadata {
+                title: Some("ZedDotDevModel".to_owned()),
+                default: Some(CloudModel::default().id().into()),
+                examples: variants.into_iter().map(Into::into).collect(),
+                ..Default::default()
+            })),
+            ..Default::default()
+        })
+    }
+}
+
+impl CloudModel {
+    pub fn id(&self) -> &str {
+        match self {
+            Self::Gpt3Point5Turbo => "gpt-3.5-turbo",
+            Self::Gpt4 => "gpt-4",
+            Self::Gpt4Turbo => "gpt-4-turbo-preview",
+            Self::Gpt4Omni => "gpt-4o",
+            Self::Gpt4OmniMini => "gpt-4o-mini",
+            Self::Claude3_5Sonnet => "claude-3-5-sonnet",
+            Self::Claude3Opus => "claude-3-opus",
+            Self::Claude3Sonnet => "claude-3-sonnet",
+            Self::Claude3Haiku => "claude-3-haiku",
+            Self::Gemini15Pro => "gemini-1.5-pro",
+            Self::Gemini15Flash => "gemini-1.5-flash",
+            Self::Custom(id) => id,
+        }
+    }
+
+    pub fn display_name(&self) -> &str {
+        match self {
+            Self::Gpt3Point5Turbo => "GPT 3.5 Turbo",
+            Self::Gpt4 => "GPT 4",
+            Self::Gpt4Turbo => "GPT 4 Turbo",
+            Self::Gpt4Omni => "GPT 4 Omni",
+            Self::Gpt4OmniMini => "GPT 4 Omni Mini",
+            Self::Claude3_5Sonnet => "Claude 3.5 Sonnet",
+            Self::Claude3Opus => "Claude 3 Opus",
+            Self::Claude3Sonnet => "Claude 3 Sonnet",
+            Self::Claude3Haiku => "Claude 3 Haiku",
+            Self::Gemini15Pro => "Gemini 1.5 Pro",
+            Self::Gemini15Flash => "Gemini 1.5 Flash",
+            Self::Custom(id) => id.as_str(),
+        }
+    }
+
+    pub fn max_token_count(&self) -> usize {
+        match self {
+            Self::Gpt3Point5Turbo => 2048,
+            Self::Gpt4 => 4096,
+            Self::Gpt4Turbo | Self::Gpt4Omni => 128000,
+            Self::Gpt4OmniMini => 128000,
+            Self::Claude3_5Sonnet
+            | Self::Claude3Opus
+            | Self::Claude3Sonnet
+            | Self::Claude3Haiku => 200000,
+            Self::Gemini15Pro => 128000,
+            Self::Gemini15Flash => 32000,
+            Self::Custom(_) => 4096, // TODO: Make this configurable
+        }
+    }
+
+    pub fn preprocess_request(&self, request: &mut LanguageModelRequest) {
+        match self {
+            Self::Claude3Opus | Self::Claude3Sonnet | Self::Claude3Haiku => {
+                request.preprocess_anthropic()
+            }
+            _ => {}
+        }
+    }
+}

crates/language_model/src/model/mod.rs

@@ -0,0 +1,60 @@
+pub mod cloud_model;
+
+pub use anthropic::Model as AnthropicModel;
+pub use cloud_model::*;
+pub use ollama::Model as OllamaModel;
+pub use open_ai::Model as OpenAiModel;
+
+use serde::{Deserialize, Serialize};
+
+#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
+pub enum LanguageModel {
+    Cloud(CloudModel),
+    OpenAi(OpenAiModel),
+    Anthropic(AnthropicModel),
+    Ollama(OllamaModel),
+}
+
+impl Default for LanguageModel {
+    fn default() -> Self {
+        LanguageModel::Cloud(CloudModel::default())
+    }
+}
+
+impl LanguageModel {
+    pub fn telemetry_id(&self) -> String {
+        match self {
+            LanguageModel::OpenAi(model) => format!("openai/{}", model.id()),
+            LanguageModel::Anthropic(model) => format!("anthropic/{}", model.id()),
+            LanguageModel::Cloud(model) => format!("zed.dev/{}", model.id()),
+            LanguageModel::Ollama(model) => format!("ollama/{}", model.id()),
+        }
+    }
+
+    pub fn display_name(&self) -> String {
+        match self {
+            LanguageModel::OpenAi(model) => model.display_name().into(),
+            LanguageModel::Anthropic(model) => model.display_name().into(),
+            LanguageModel::Cloud(model) => model.display_name().into(),
+            LanguageModel::Ollama(model) => model.display_name().into(),
+        }
+    }
+
+    pub fn max_token_count(&self) -> usize {
+        match self {
+            LanguageModel::OpenAi(model) => model.max_token_count(),
+            LanguageModel::Anthropic(model) => model.max_token_count(),
+            LanguageModel::Cloud(model) => model.max_token_count(),
+            LanguageModel::Ollama(model) => model.max_token_count(),
+        }
+    }
+
+    pub fn id(&self) -> &str {
+        match self {
+            LanguageModel::OpenAi(model) => model.id(),
+            LanguageModel::Anthropic(model) => model.id(),
+            LanguageModel::Cloud(model) => model.id(),
+            LanguageModel::Ollama(model) => model.id(),
+        }
+    }
+}

crates/language_model/src/request.rs

@@ -0,0 +1,110 @@
+use crate::{
+    model::{CloudModel, LanguageModel},
+    role::Role,
+};
+use serde::{Deserialize, Serialize};
+
+#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
+pub struct LanguageModelRequestMessage {
+    pub role: Role,
+    pub content: String,
+}
+
+impl LanguageModelRequestMessage {
+    pub fn to_proto(&self) -> proto::LanguageModelRequestMessage {
+        proto::LanguageModelRequestMessage {
+            role: self.role.to_proto() as i32,
+            content: self.content.clone(),
+            tool_calls: Vec::new(),
+            tool_call_id: None,
+        }
+    }
+}
+
+#[derive(Debug, Default, Serialize, Deserialize)]
+pub struct LanguageModelRequest {
+    pub model: LanguageModel,
+    pub messages: Vec<LanguageModelRequestMessage>,
+    pub stop: Vec<String>,
+    pub temperature: f32,
+}
+
+impl LanguageModelRequest {
+    pub fn to_proto(&self) -> proto::CompleteWithLanguageModel {
+        proto::CompleteWithLanguageModel {
+            model: self.model.id().to_string(),
+            messages: self.messages.iter().map(|m| m.to_proto()).collect(),
+            stop: self.stop.clone(),
+            temperature: self.temperature,
+            tool_choice: None,
+            tools: Vec::new(),
+        }
+    }
+
+    /// Before we send the request to the server, we can perform fixups on it appropriate to the model.
+    pub fn preprocess(&mut self) {
+        match &self.model {
+            LanguageModel::OpenAi(_) => {}
+            LanguageModel::Anthropic(_) => {}
+            LanguageModel::Ollama(_) => {}
+            LanguageModel::Cloud(model) => match model {
+                CloudModel::Claude3Opus
+                | CloudModel::Claude3Sonnet
+                | CloudModel::Claude3Haiku
+                | CloudModel::Claude3_5Sonnet => {
+                    self.preprocess_anthropic();
+                }
+                _ => {}
+            },
+        }
+    }
+
+    pub fn preprocess_anthropic(&mut self) {
+        let mut new_messages: Vec<LanguageModelRequestMessage> = Vec::new();
+        let mut system_message = String::new();
+
+        for message in self.messages.drain(..) {
+            if message.content.is_empty() {
+                continue;
+            }
+
+            match message.role {
+                Role::User | Role::Assistant => {
+                    if let Some(last_message) = new_messages.last_mut() {
+                        if last_message.role == message.role {
+                            last_message.content.push_str("\n\n");
+                            last_message.content.push_str(&message.content);
+                            continue;
+                        }
+                    }
+
+                    new_messages.push(message);
+                }
+                Role::System => {
+                    if !system_message.is_empty() {
+                        system_message.push_str("\n\n");
+                    }
+                    system_message.push_str(&message.content);
+                }
+            }
+        }
+
+        if !system_message.is_empty() {
+            new_messages.insert(
+                0,
+                LanguageModelRequestMessage {
+                    role: Role::System,
+                    content: system_message,
+                },
+            );
+        }
+
+        self.messages = new_messages;
+    }
+}
+
+#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
+pub struct LanguageModelResponseMessage {
+    pub role: Option<Role>,
+    pub content: Option<String>,
+}

crates/language_model/src/role.rs

@@ -0,0 +1,68 @@
+use serde::{Deserialize, Serialize};
+use std::fmt::{self, Display};
+
+#[derive(Clone, Copy, Serialize, Deserialize, Debug, Eq, PartialEq)]
+#[serde(rename_all = "lowercase")]
+pub enum Role {
+    User,
+    Assistant,
+    System,
+}
+
+impl Role {
+    pub fn from_proto(role: i32) -> Role {
+        match proto::LanguageModelRole::from_i32(role) {
+            Some(proto::LanguageModelRole::LanguageModelUser) => Role::User,
+            Some(proto::LanguageModelRole::LanguageModelAssistant) => Role::Assistant,
+            Some(proto::LanguageModelRole::LanguageModelSystem) => Role::System,
+            Some(proto::LanguageModelRole::LanguageModelTool) => Role::System,
+            None => Role::User,
+        }
+    }
+
+    pub fn to_proto(&self) -> proto::LanguageModelRole {
+        match self {
+            Role::User => proto::LanguageModelRole::LanguageModelUser,
+            Role::Assistant => proto::LanguageModelRole::LanguageModelAssistant,
+            Role::System => proto::LanguageModelRole::LanguageModelSystem,
+        }
+    }
+
+    pub fn cycle(self) -> Role {
+        match self {
+            Role::User => Role::Assistant,
+            Role::Assistant => Role::System,
+            Role::System => Role::User,
+        }
+    }
+}
+
+impl Display for Role {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            Role::User => write!(f, "user"),
+            Role::Assistant => write!(f, "assistant"),
+            Role::System => write!(f, "system"),
+        }
+    }
+}
+
+impl From<Role> for ollama::Role {
+    fn from(val: Role) -> Self {
+        match val {
+            Role::User => ollama::Role::User,
+            Role::Assistant => ollama::Role::Assistant,
+            Role::System => ollama::Role::System,
+        }
+    }
+}
+
+impl From<Role> for open_ai::Role {
+    fn from(val: Role) -> Self {
+        match val {
+            Role::User => open_ai::Role::User,
+            Role::Assistant => open_ai::Role::Assistant,
+            Role::System => open_ai::Role::System,
+        }
+    }
+}