From 3d5ddcccf08ea4d421b11474cf1190fc6bb1b4a1 Mon Sep 17 00:00:00 2001
From: Bennet Bo Fenner <bennet@zed.dev>
Date: Fri, 10 Oct 2025 14:59:52 +0200
Subject: [PATCH] ollama: Resolve context window size via API (#39941)

Previously we were guessing the context window size here:
https://github.com/zed-industries/zed/blob/8c3f09e31e3588a2494042dfe77bb3dccab1f7ba/crates/ollama/src/ollama.rs#L22

This is inaccurate and must be updated manually. This PR ensures that we
extract the context window size from the request in the same way that
the Ollama CLI does when running `ollama show <model-name>` (Relevant
code is
[here](https://github.com/ollama/ollama/blob/3d32249c749c6f77c1dc8a7cb55ae74fc2f4c08b/cmd/cmd.go#L860))

The format looks like this:

```json
{
  "model_info": {
    "general.architecture": "llama",
    "llama.context_length": 132000
  }
}
```

Once this PR is merged we could technically remove the old code
https://github.com/zed-industries/zed/blob/8c3f09e31e3588a2494042dfe77bb3dccab1f7ba/crates/ollama/src/ollama.rs#L22
I decided to keep it for now, as it is unclear if the necessary fields
are available via the API on older Ollama versions.

Release Notes:

- Fixed an issue where Ollama models would use the wrong context window
size
---
 crates/language_models/src/provider/ollama.rs | 10 +--
 crates/ollama/src/ollama.rs                   | 71 ++++++++++++++++++-
 2 files changed, 74 insertions(+), 7 deletions(-)

diff --git a/crates/language_models/src/provider/ollama.rs b/crates/language_models/src/provider/ollama.rs
index d3a4289f24f651507ed493af228ac30314d975d6..a25ecbe01aa659817b41cac54d76871b4742ea66 100644
--- a/crates/language_models/src/provider/ollama.rs
+++ b/crates/language_models/src/provider/ollama.rs
@@ -119,16 +119,16 @@ impl State {
                     let api_key = api_key.clone();
                     async move {
                         let name = model.name.as_str();
-                        let capabilities =
+                        let model =
                             show_model(http_client.as_ref(), &api_url, api_key.as_deref(), name)
                                 .await?;
                         let ollama_model = ollama::Model::new(
                             name,
                             None,
-                            None,
-                            Some(capabilities.supports_tools()),
-                            Some(capabilities.supports_vision()),
-                            Some(capabilities.supports_thinking()),
+                            model.context_length,
+                            Some(model.supports_tools()),
+                            Some(model.supports_vision()),
+                            Some(model.supports_thinking()),
                         );
                         Ok(ollama_model)
                     }
diff --git a/crates/ollama/src/ollama.rs b/crates/ollama/src/ollama.rs
index fe0b609646eb2e20090e55b1e6b6ed02fb571776..0ed3b6da17d952cc874485337ec380ef3ca990a8 100644
--- a/crates/ollama/src/ollama.rs
+++ b/crates/ollama/src/ollama.rs
@@ -189,10 +189,74 @@ pub struct ModelDetails {
     pub quantization_level: String,
 }
 
-#[derive(Deserialize, Debug)]
+#[derive(Debug)]
 pub struct ModelShow {
-    #[serde(default)]
     pub capabilities: Vec<String>,
+    pub context_length: Option<u64>,
+    pub architecture: Option<String>,
+}
+
+impl<'de> Deserialize<'de> for ModelShow {
+    fn deserialize<D>(deserializer: D) -> std::result::Result<Self, D::Error>
+    where
+        D: serde::Deserializer<'de>,
+    {
+        use serde::de::{self, MapAccess, Visitor};
+        use std::fmt;
+
+        struct ModelShowVisitor;
+
+        impl<'de> Visitor<'de> for ModelShowVisitor {
+            type Value = ModelShow;
+
+            fn expecting(&self, formatter: &mut fmt::Formatter) -> fmt::Result {
+                formatter.write_str("a ModelShow object")
+            }
+
+            fn visit_map<A>(self, mut map: A) -> Result<Self::Value, A::Error>
+            where
+                A: MapAccess<'de>,
+            {
+                let mut capabilities: Vec<String> = Vec::new();
+                let mut architecture: Option<String> = None;
+                let mut context_length: Option<u64> = None;
+
+                while let Some(key) = map.next_key::<String>()? {
+                    match key.as_str() {
+                        "capabilities" => {
+                            capabilities = map.next_value()?;
+                        }
+                        "model_info" => {
+                            let model_info: Value = map.next_value()?;
+                            if let Value::Object(obj) = model_info {
+                                architecture = obj
+                                    .get("general.architecture")
+                                    .and_then(|v| v.as_str())
+                                    .map(String::from);
+
+                                if let Some(arch) = &architecture {
+                                    context_length = obj
+                                        .get(&format!("{}.context_length", arch))
+                                        .and_then(|v| v.as_u64());
+                                }
+                            }
+                        }
+                        _ => {
+                            let _: de::IgnoredAny = map.next_value()?;
+                        }
+                    }
+                }
+
+                Ok(ModelShow {
+                    capabilities,
+                    context_length,
+                    architecture,
+                })
+            }
+        }
+
+        deserializer.deserialize_map(ModelShowVisitor)
+    }
 }
 
 impl ModelShow {
@@ -470,6 +534,9 @@ mod tests {
         assert!(result.supports_tools());
         assert!(result.capabilities.contains(&"tools".to_string()));
         assert!(result.capabilities.contains(&"completion".to_string()));
+
+        assert_eq!(result.architecture, Some("llama".to_string()));
+        assert_eq!(result.context_length, Some(131072));
     }
 
     #[test]