Back to Agno

TEST_LOG - Gemini Interactions API Cookbooks

cookbook/90_models/google/gemini_interactions/TEST_LOG.md

2.6.96.6 KB
Original Source

TEST_LOG - Gemini Interactions API Cookbooks

Test Environment

  • Python: 3.12
  • SDK: google-genai (latest)
  • Model: gemini-3.5-flash
  • Date: 2026-05-14

basic.py

Status: PASS

Description: Tests all four invocation modes: sync, sync+streaming, async, async+streaming. All modes returned coherent text responses.

Result: All four modes produced valid 2-sentence horror stories. Async mode showed expected aiohttp fallback warning.


tool_use.py

Status: PASS

Description: Tests function calling with WebSearchTools. Agent correctly invokes the web search tool and returns a synthesized answer.

Result: Sync, streaming, and async streaming all produced detailed current events responses about France.


multi_turn.py

Status: PASS

Description: Tests multi-turn conversation with server-side history via previous_interaction_id. Agent maintains context across turns.

Result: Agent correctly referenced prior conversation context (hiking interest) in follow-up responses.


thinking.py

Status: PASS

Description: Tests thinking/reasoning mode with thinking_budget configuration.

Result: Model produced detailed mathematical reasoning about triangle angles with multiple proof approaches.


search.py

Status: PASS

Description: Tests built-in Google Search tool (search=True). Agent uses Google Search grounding for real-time information.

Result: Returned current news with citations including geopolitics, tech, sports, and entertainment.


image_understanding.py

Status: PASS

Description: Tests image understanding from URL. Image is downloaded via httpx and sent as base64 data.

Result: Model correctly identified and described a black Labrador puppy on wooden floorboards with detailed analysis of composition, lighting, and mood.


audio_understanding.py

Status: PASS

Description: Tests audio understanding from URL. Audio file is downloaded and sent as base64 data.

Result: Model identified and described the audio clip contents correctly.


video_understanding.py

Status: PASS

Description: Tests video understanding from URL. Video file is downloaded and sent as base64 data.

Result: Model described the video scene in detail including setting, background elements, people, and atmosphere.


document_processing.py

Status: PASS

Description: Tests PDF document processing from URL (arxiv "Attention Is All You Need" paper).

Result: Model produced a comprehensive summary of the Transformer paper including key innovations, performance results, and significance.


structured_output.py

Status: PASS

Description: Tests structured output with Pydantic schema (MovieReview). Uses output_schema parameter with response_format using TextResponseFormatParam.

Result: Returned a properly typed MovieReview object with title="The Matrix", year=1999, genre="Sci-Fi", rating=8.7, and a coherent summary.


image_generation.py

Status: FAIL (API limitation)

Description: Tests image generation using response_modalities=["text", "image"]. The Interactions API does not currently support image generation output with gemini-3.5-flash.

Result: Model returned text instead of generating an image. Image generation is not yet supported through the Interactions API. The cookbook is kept as a reference for when support is added.


Sanity sweep for PR #8045 (2026-05-21)

Branch fix/gemini-interactions-agent-tool-calls at commit ab12588c2. SDK upgraded to google-genai 2.5.0 in .venvs/demo to satisfy the public google.genai.interactions import path the PR uses.

basic.py

Status: PASS

Description: All four invocation modes (sync, sync+stream, async, async+stream) against gemini-3.5-flash.

Result: All modes produced coherent 2-sentence horror stories. No errors.


multi_turn.py

Status: PASS

Description: Three sequential turns sharing context via previous_interaction_id (SqliteDb persistence).

Result: Each turn correctly built on the previous one (e.g. the third-turn hike recommendation reflected the user profile established earlier). No 400s, no stale-state issues.


tool_use.py

Status: PASS

Description: Client-declared WebSearchTools exercised across sync, sync+stream, async+stream. Regression check for the Codex P1 fix (FunctionCallStep with no matching FunctionResultStep should fall through to client dispatch).

Result: WebSearch tool was correctly dispatched locally in all three modes and produced a synthesized France news answer. No "Function not found" errors.


structured_output.py

Status: PASS

Description: Pydantic MovieReview output_schema via the Interactions API's TextResponseFormatParam.

Result: Returned a fully populated MovieReview (title="The Matrix", year=1999, genre="Sci-Fi", rating=8.7) parsed correctly.


thinking.py

Status: PASS

Description: thinking_level="high" reasoning across sync and async+stream.

Result: Both modes returned coherent geometry explanations. reasoning_content flowed through the parser (verified via the <reasoning> debug block).


search.py

Status: PASS

Description: Built-in search=True tool across sync and async+stream.

Result: Both modes returned current-news answers grounded in Google search; no parser errors.


antigravity.py

Status: PASS

Description: Basic non-streaming agent path call ("What is the capital of France?").

Result: Returned "Paris" in 11.1s. No errors. Sandbox provisioned and tool calls completed server-side without leaking to local dispatch.


antigravity_streaming.py

Status: PASS

Description: Agent path with stream=True — the primary case the PR fixes. Asks the model to fetch Hacker News + summarize 5 stories AND run an async stream finding three trending Python repos.

Result: Both runs completed without errors. 26 Server-side tool call: / Server-side reasoning: log lines emitted across the run. Final markdown summaries rendered cleanly. No "Function list_files not found" or 400 invalid_request from the original bug.


antigravity_multi_turn.py

Status: PASS

Description: Three chained Antigravity turns through SqliteDb (plot solar growth → embed in HTML deck → revise deck).

Result: All three turns completed without errors. Second and third turns sent previous_interaction_id (confirmed via 2 occurrences in the debug log) and the sandbox state carried forward. Each turn's ToolExecution records (list_files, code_execution, browser ops, etc.) populated run_response.tools.