llm-gemini 0.13.1__tar.gz → 0.14__tar.gz

{llm_gemini-0.13.1 → llm_gemini-0.14}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: llm-gemini
-Version: 0.13.1
+Version: 0.14
 Summary: LLM plugin to access Google's Gemini family of models
 Author: Simon Willison
 License: Apache-2.0
@@ -17,6 +17,7 @@ Requires-Dist: ijson
 Provides-Extra: test
 Requires-Dist: pytest; extra == "test"
 Requires-Dist: pytest-recording; extra == "test"
+Requires-Dist: pytest-asyncio; extra == "test"
 Requires-Dist: nest-asyncio; extra == "test"
 
 # llm-gemini
@@ -145,7 +146,7 @@ llm chat -m gemini-1.5-pro-latest
 
 ## Embeddings
 
-The plugin also adds support for the `text-embedding-004` embedding model.
+The plugin also adds support for the `gemini-embedding-exp-03-07` and `text-embedding-004` embedding models.
 
 Run that against a single string like this:
 ```bash
@@ -153,10 +154,20 @@ llm embed -m text-embedding-004 -c 'hello world'
 ```
 This returns a JSON array of 768 numbers.
 
+The `gemini-embedding-exp-03-07` model is larger, returning 3072 numbers. You can also use variants of it that are truncated down to smaller sizes:
+
+- `gemini-embedding-exp-03-07` - 3072 numbers
+- `gemini-embedding-exp-03-07-2048` - 2048 numbers
+- `gemini-embedding-exp-03-07-1024` - 1024 numbers
+- `gemini-embedding-exp-03-07-512` - 512 numbers
+- `gemini-embedding-exp-03-07-256` - 256 numbers
+- `gemini-embedding-exp-03-07-128` - 128 numbers
+
 This command will embed every `README.md` file in child directories of the current directory and store the results in a SQLite database called `embed.db` in a collection called `readmes`:
 
 ```bash
-llm embed-multi readmes --files . '*/README.md' -d embed.db -m text-embedding-004
+llm embed-multi readmes -d embed.db -m gemini-embedding-exp-03-07-128 \
+  --files . '*/README.md'
 ```
 You can then run similarity searches against that collection like this:
 ```bash

{llm_gemini-0.13.1 → llm_gemini-0.14}/README.md

@@ -124,7 +124,7 @@ llm chat -m gemini-1.5-pro-latest
 
 ## Embeddings
 
-The plugin also adds support for the `text-embedding-004` embedding model.
+The plugin also adds support for the `gemini-embedding-exp-03-07` and `text-embedding-004` embedding models.
 
 Run that against a single string like this:
 ```bash
@@ -132,10 +132,20 @@ llm embed -m text-embedding-004 -c 'hello world'
 ```
 This returns a JSON array of 768 numbers.
 
+The `gemini-embedding-exp-03-07` model is larger, returning 3072 numbers. You can also use variants of it that are truncated down to smaller sizes:
+
+- `gemini-embedding-exp-03-07` - 3072 numbers
+- `gemini-embedding-exp-03-07-2048` - 2048 numbers
+- `gemini-embedding-exp-03-07-1024` - 1024 numbers
+- `gemini-embedding-exp-03-07-512` - 512 numbers
+- `gemini-embedding-exp-03-07-256` - 256 numbers
+- `gemini-embedding-exp-03-07-128` - 128 numbers
+
 This command will embed every `README.md` file in child directories of the current directory and store the results in a SQLite database called `embed.db` in a collection called `readmes`:
 
 ```bash
-llm embed-multi readmes --files . '*/README.md' -d embed.db -m text-embedding-004
+llm embed-multi readmes -d embed.db -m gemini-embedding-exp-03-07-128 \
+  --files . '*/README.md'
 ```
 You can then run similarity searches against that collection like this:
 ```bash

{llm_gemini-0.13.1 → llm_gemini-0.14}/llm_gemini.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: llm-gemini
-Version: 0.13.1
+Version: 0.14
 Summary: LLM plugin to access Google's Gemini family of models
 Author: Simon Willison
 License: Apache-2.0
@@ -17,6 +17,7 @@ Requires-Dist: ijson
 Provides-Extra: test
 Requires-Dist: pytest; extra == "test"
 Requires-Dist: pytest-recording; extra == "test"
+Requires-Dist: pytest-asyncio; extra == "test"
 Requires-Dist: nest-asyncio; extra == "test"
 
 # llm-gemini
@@ -145,7 +146,7 @@ llm chat -m gemini-1.5-pro-latest
 
 ## Embeddings
 
-The plugin also adds support for the `text-embedding-004` embedding model.
+The plugin also adds support for the `gemini-embedding-exp-03-07` and `text-embedding-004` embedding models.
 
 Run that against a single string like this:
 ```bash
@@ -153,10 +154,20 @@ llm embed -m text-embedding-004 -c 'hello world'
 ```
 This returns a JSON array of 768 numbers.
 
+The `gemini-embedding-exp-03-07` model is larger, returning 3072 numbers. You can also use variants of it that are truncated down to smaller sizes:
+
+- `gemini-embedding-exp-03-07` - 3072 numbers
+- `gemini-embedding-exp-03-07-2048` - 2048 numbers
+- `gemini-embedding-exp-03-07-1024` - 1024 numbers
+- `gemini-embedding-exp-03-07-512` - 512 numbers
+- `gemini-embedding-exp-03-07-256` - 256 numbers
+- `gemini-embedding-exp-03-07-128` - 128 numbers
+
 This command will embed every `README.md` file in child directories of the current directory and store the results in a SQLite database called `embed.db` in a collection called `readmes`:
 
 ```bash
-llm embed-multi readmes --files . '*/README.md' -d embed.db -m text-embedding-004
+llm embed-multi readmes -d embed.db -m gemini-embedding-exp-03-07-128 \
+  --files . '*/README.md'
 ```
 You can then run similarity searches against that collection like this:
 ```bash

{llm_gemini-0.13.1 → llm_gemini-0.14}/llm_gemini.egg-info/requires.txt

@@ -5,4 +5,5 @@ ijson
 [test]
 pytest
 pytest-recording
+pytest-asyncio
 nest-asyncio

{llm_gemini-0.13.1 → llm_gemini-0.14}/llm_gemini.py

@@ -378,9 +378,19 @@ class AsyncGeminiPro(_SharedGemini, llm.AsyncKeyModel):
 
 @llm.hookimpl
 def register_embedding_models(register):
+    register(GeminiEmbeddingModel("text-embedding-004", "text-embedding-004"))
+    # gemini-embedding-exp-03-07 in different truncation sizes
     register(
-        GeminiEmbeddingModel("text-embedding-004", "text-embedding-004"),
+        GeminiEmbeddingModel(
+            "gemini-embedding-exp-03-07", "gemini-embedding-exp-03-07"
+        ),
     )
+    for i in (128, 256, 512, 1024, 2048):
+        register(
+            GeminiEmbeddingModel(
+                f"gemini-embedding-exp-03-07-{i}", f"gemini-embedding-exp-03-07", i
+            ),
+        )
 
 
 class GeminiEmbeddingModel(llm.EmbeddingModel):
@@ -388,9 +398,10 @@ class GeminiEmbeddingModel(llm.EmbeddingModel):
     key_env_var = "LLM_GEMINI_KEY"
     batch_size = 20
 
-    def __init__(self, model_id, gemini_model_id):
+    def __init__(self, model_id, gemini_model_id, truncate=None):
         self.model_id = model_id
         self.gemini_model_id = gemini_model_id
+        self.truncate = truncate
 
     def embed_batch(self, items):
         headers = {
@@ -416,4 +427,7 @@ class GeminiEmbeddingModel(llm.EmbeddingModel):
             )
 
         response.raise_for_status()
-        return [item["values"] for item in response.json()["embeddings"]]
+        values = [item["values"] for item in response.json()["embeddings"]]
+        if self.truncate:
+            values = [value[: self.truncate] for value in values]
+        return values

{llm_gemini-0.13.1 → llm_gemini-0.14}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "llm-gemini"
-version = "0.13.1"
+version = "0.14"
 description = "LLM plugin to access Google's Gemini family of models"
 readme = "README.md"
 authors = [{name = "Simon Willison"}]
@@ -24,4 +24,4 @@ CI = "https://github.com/simonw/llm-gemini/actions"
 gemini = "llm_gemini"
 
 [project.optional-dependencies]
-test = ["pytest", "pytest-recording", "nest-asyncio"]
+test = ["pytest", "pytest-recording", "pytest-asyncio", "nest-asyncio"]

{llm_gemini-0.13.1 → llm_gemini-0.14}/tests/test_gemini.py

@@ -102,3 +102,24 @@ async def test_prompt_with_pydantic_schema():
         "modelVersion": "gemini-1.5-flash-latest",
     }
     assert response.input_tokens == 10
+
+
+@pytest.mark.vcr
+@pytest.mark.parametrize(
+    "model_id",
+    (
+        "gemini-embedding-exp-03-07",
+        "gemini-embedding-exp-03-07-128",
+        "gemini-embedding-exp-03-07-512",
+    ),
+)
+def test_embedding(model_id, monkeypatch):
+    monkeypatch.setenv("LLM_GEMINI_KEY", GEMINI_API_KEY)
+    model = llm.get_embedding_model(model_id)
+    response = model.embed("Some text goes here")
+    expected_length = 3072
+    if model_id.endswith("-128"):
+        expected_length = 128
+    elif model_id.endswith("-512"):
+        expected_length = 512
+    assert len(response) == expected_length