llm-gemini 0.13a0__tar.gz → 0.14__tar.gz

{llm_gemini-0.13a0 → llm_gemini-0.14}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: llm-gemini
-Version: 0.13a0
+Version: 0.14
 Summary: LLM plugin to access Google's Gemini family of models
 Author: Simon Willison
 License: Apache-2.0
@@ -11,12 +11,13 @@ Project-URL: CI, https://github.com/simonw/llm-gemini/actions
 Classifier: License :: OSI Approved :: Apache Software License
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: llm>=0.23a0
+Requires-Dist: llm>=0.23
 Requires-Dist: httpx
 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-0.13a0 → 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-0.13a0 → llm_gemini-0.14}/llm_gemini.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: llm-gemini
-Version: 0.13a0
+Version: 0.14
 Summary: LLM plugin to access Google's Gemini family of models
 Author: Simon Willison
 License: Apache-2.0
@@ -11,12 +11,13 @@ Project-URL: CI, https://github.com/simonw/llm-gemini/actions
 Classifier: License :: OSI Approved :: Apache Software License
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: llm>=0.23a0
+Requires-Dist: llm>=0.23
 Requires-Dist: httpx
 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-0.13a0 → llm_gemini-0.14}/llm_gemini.egg-info/requires.txt

@@ -1,8 +1,9 @@
-llm>=0.23a0
+llm>=0.23
 httpx
 ijson
 
 [test]
 pytest
 pytest-recording
+pytest-asyncio
 nest-asyncio

{llm_gemini-0.13a0 → llm_gemini-0.14}/llm_gemini.py

@@ -65,8 +65,16 @@ def register_models(register):
     ]:
         can_google_search = model_id in GOOGLE_SEARCH_MODELS
         register(
-            GeminiPro(model_id, can_google_search=can_google_search),
-            AsyncGeminiPro(model_id, can_google_search=can_google_search),
+            GeminiPro(
+                model_id,
+                can_google_search=can_google_search,
+                can_schema="flash-thinking" not in model_id,
+            ),
+            AsyncGeminiPro(
+                model_id,
+                can_google_search=can_google_search,
+                can_schema="flash-thinking" not in model_id,
+            ),
         )
 
 
@@ -82,7 +90,7 @@ def resolve_type(attachment):
 
 def cleanup_schema(schema):
     "Gemini supports only a subset of JSON schema"
-    keys_to_remove = ("$schema", "additionalProperties")
+    keys_to_remove = ("$schema", "additionalProperties", "title")
     # Recursively remove them
     if isinstance(schema, dict):
         for key in keys_to_remove:
@@ -186,9 +194,10 @@ class _SharedGemini:
             default=None,
         )
 
-    def __init__(self, model_id, can_google_search=False):
+    def __init__(self, model_id, can_google_search=False, can_schema=False):
         self.model_id = model_id
         self.can_google_search = can_google_search
+        self.supports_schema = can_schema
         if can_google_search:
             self.Options = self.OptionsWithGoogleSearch
 
@@ -279,9 +288,17 @@ class _SharedGemini:
             return f'```\n{part["codeExecutionResult"]["output"].strip()}\n```\n'
         return ""
 
+    def process_candidates(self, candidates):
+        # We only use the first candidate
+        for part in candidates[0]["content"]["parts"]:
+            yield self.process_part(part)
+
     def set_usage(self, response):
         try:
-            usage = response.response_json[-1].pop("usageMetadata")
+            # Don't record the "content" key from that last candidate
+            for candidate in response.response_json["candidates"]:
+                candidate.pop("content", None)
+            usage = response.response_json.pop("usageMetadata")
             input_tokens = usage.pop("promptTokenCount", None)
             output_tokens = usage.pop("candidatesTokenCount", None)
             usage.pop("totalTokenCount", None)
@@ -311,17 +328,16 @@ class GeminiPro(_SharedGemini, llm.KeyModel):
             for chunk in http_response.iter_bytes():
                 coro.send(chunk)
                 if events:
-                    event = events[0]
-                    if isinstance(event, dict) and "error" in event:
-                        raise llm.ModelError(event["error"]["message"])
-                    try:
-                        part = event["candidates"][0]["content"]["parts"][0]
-                        yield self.process_part(part)
-                    except KeyError:
-                        yield ""
-                    gathered.append(event)
+                    for event in events:
+                        if isinstance(event, dict) and "error" in event:
+                            raise llm.ModelError(event["error"]["message"])
+                        try:
+                            yield from self.process_candidates(event["candidates"])
+                        except KeyError:
+                            yield ""
+                        gathered.append(event)
                     events.clear()
-        response.response_json = gathered
+        response.response_json = gathered[-1]
         self.set_usage(response)
 
 
@@ -344,25 +360,37 @@ class AsyncGeminiPro(_SharedGemini, llm.AsyncKeyModel):
                 async for chunk in http_response.aiter_bytes():
                     coro.send(chunk)
                     if events:
-                        event = events[0]
-                        if isinstance(event, dict) and "error" in event:
-                            raise llm.ModelError(event["error"]["message"])
-                        try:
-                            part = event["candidates"][0]["content"]["parts"][0]
-                            yield self.process_part(part)
-                        except KeyError:
-                            yield ""
-                        gathered.append(event)
+                        for event in events:
+                            if isinstance(event, dict) and "error" in event:
+                                raise llm.ModelError(event["error"]["message"])
+                            try:
+                                for chunk in self.process_candidates(
+                                    event["candidates"]
+                                ):
+                                    yield chunk
+                            except KeyError:
+                                yield ""
+                            gathered.append(event)
                         events.clear()
-        response.response_json = gathered
+        response.response_json = gathered[-1]
         self.set_usage(response)
 
 
 @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):
@@ -370,9 +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 = {
@@ -398,4 +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-0.13a0 → llm_gemini-0.14}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "llm-gemini"
-version = "0.13a0"
+version = "0.14"
 description = "LLM plugin to access Google's Gemini family of models"
 readme = "README.md"
 authors = [{name = "Simon Willison"}]
@@ -9,7 +9,7 @@ classifiers = [
     "License :: OSI Approved :: Apache Software License"
 ]
 dependencies = [
-    "llm>=0.23a0",
+    "llm>=0.23",
     "httpx",
     "ijson"
 ]
@@ -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-0.14/tests/test_gemini.py

@@ -0,0 +1,125 @@
+import llm
+import nest_asyncio
+import json
+import os
+import pytest
+import pydantic
+
+nest_asyncio.apply()
+
+GEMINI_API_KEY = os.environ.get("PYTEST_GEMINI_API_KEY", None) or "gm-..."
+
+
+@pytest.mark.vcr
+@pytest.mark.asyncio
+async def test_prompt():
+    model = llm.get_model("gemini-1.5-flash-latest")
+    response = model.prompt("Name for a pet pelican, just the name", key=GEMINI_API_KEY)
+    assert str(response) == "Percy\n"
+    assert response.response_json == {
+        "candidates": [
+            {
+                "finishReason": "STOP",
+                "safetyRatings": [
+                    {
+                        "category": "HARM_CATEGORY_HATE_SPEECH",
+                        "probability": "NEGLIGIBLE",
+                    },
+                    {
+                        "category": "HARM_CATEGORY_DANGEROUS_CONTENT",
+                        "probability": "NEGLIGIBLE",
+                    },
+                    {
+                        "category": "HARM_CATEGORY_HARASSMENT",
+                        "probability": "NEGLIGIBLE",
+                    },
+                    {
+                        "category": "HARM_CATEGORY_SEXUALLY_EXPLICIT",
+                        "probability": "NEGLIGIBLE",
+                    },
+                ],
+            }
+        ],
+        "modelVersion": "gemini-1.5-flash-latest",
+    }
+    assert response.token_details == {
+        "promptTokensDetails": [{"modality": "TEXT", "tokenCount": 9}],
+        "candidatesTokensDetails": [{"modality": "TEXT", "tokenCount": 2}],
+    }
+    assert response.input_tokens == 9
+    assert response.output_tokens == 2
+
+    # And try it async too
+    async_model = llm.get_async_model("gemini-1.5-flash-latest")
+    response = await async_model.prompt(
+        "Name for a pet pelican, just the name", key=GEMINI_API_KEY
+    )
+    text = await response.text()
+    assert text == "Percy\n"
+
+
+@pytest.mark.vcr
+@pytest.mark.asyncio
+async def test_prompt_with_pydantic_schema():
+    class Dog(pydantic.BaseModel):
+        name: str
+        age: int
+        bio: str
+
+    model = llm.get_model("gemini-1.5-flash-latest")
+    response = model.prompt(
+        "Invent a cool dog", key=GEMINI_API_KEY, schema=Dog, stream=False
+    )
+    assert json.loads(response.text()) == {
+        "age": 3,
+        "bio": "A fluffy Samoyed with exceptional intelligence and a love for belly rubs. He's mastered several tricks, including fetching the newspaper and opening doors.",
+        "name": "Cloud",
+    }
+    assert response.response_json == {
+        "candidates": [
+            {
+                "finishReason": "STOP",
+                "safetyRatings": [
+                    {
+                        "category": "HARM_CATEGORY_HATE_SPEECH",
+                        "probability": "NEGLIGIBLE",
+                    },
+                    {
+                        "category": "HARM_CATEGORY_DANGEROUS_CONTENT",
+                        "probability": "NEGLIGIBLE",
+                    },
+                    {
+                        "category": "HARM_CATEGORY_HARASSMENT",
+                        "probability": "NEGLIGIBLE",
+                    },
+                    {
+                        "category": "HARM_CATEGORY_SEXUALLY_EXPLICIT",
+                        "probability": "NEGLIGIBLE",
+                    },
+                ],
+            }
+        ],
+        "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

llm_gemini-0.13a0/tests/test_gemini.py

@@ -1,37 +0,0 @@
-import llm
-import nest_asyncio
-import os
-import pytest
-
-nest_asyncio.apply()
-
-GEMINI_API_KEY = os.environ.get("PYTEST_GEMINI_API_KEY", None) or "gm-..."
-
-
-@pytest.mark.vcr
-@pytest.mark.asyncio
-async def test_prompt():
-    model = llm.get_model("gemini-1.5-flash-latest")
-    response = model.prompt("Name for a pet pelican, just the name", key=GEMINI_API_KEY)
-    assert str(response) == "Percy"
-    assert response.response_json == [
-        {
-            "candidates": [
-                {"content": {"parts": [{"text": "Percy"}], "role": "model"}}
-            ],
-            "modelVersion": "gemini-1.5-flash-latest",
-        }
-    ]
-    assert response.token_details is None
-    assert response.input_tokens == 10
-    # Not sure why our pytest-recording setup doesn't report output tokens
-    # https://github.com/simonw/llm-gemini/issues/25#issuecomment-2487464339
-    assert response.output_tokens is None
-
-    # And try it async too
-    async_model = llm.get_async_model("gemini-1.5-flash-latest")
-    response = await async_model.prompt(
-        "Name for a pet pelican, just the name", key=GEMINI_API_KEY
-    )
-    text = await response.text()
-    assert text == "Percy"