Skip to content

AI flight analysis (Gemini)

The AI Flight Analysis block in the Streamlit app (views/ai_analysis.py) calls Google Gemini to turn already-computed mission metrics into a short text report or a custom question-and-answer reply. It does not send the raw .bin log or full time series—only the structured metrics dict from compute_metrics().

What runs where

Piece Role
views/ai_analysis.py Streamlit UI: API key, mode picker, button, result display, session cache
service/ai/flight_analysis.py Builds the metrics text block, picks system prompts, calls the Gemini API
google-genai (import google.genai) Python client (see requirements.txt)

The legacy Dash app (drone_dashboard.py) does not include this feature.

API key

Set GEMINI_API_KEY in .streamlit/secrets.toml at the project root. The app also reads GEMINI_API_KEY from the process environment (app.py calls load_dotenv(), so a project-root .env may be used if you prefer). If unset, the UI shows an expander where you can paste a key for the current session (password field). Keys are obtained from Google AI Studio.

Setup is described in Run the service — step 4.

Model

The service uses a single model id:

  • gemini-3-flash-preview (_MODEL_NAME in service/ai/flight_analysis.py)

Google may rename or retire preview models; if generation fails after an API change, update _MODEL_NAME to a current Gemini model your key supports.

Analysis modes

UI label Internal mode Behaviour
Short Report short Brief summary (system prompt caps at ~150 words): key stats, optional energy line, up to two safety notes
Detailed Report detailed Full structured report: overview, performance, energy, GPS quality, vibration/mechanical comments from gyro warnings, safety, recommendations
Custom Question custom Your question in a text area, plus the same metrics context block

System instructions ask for professional, concise Markdown, bullet points inside sections, and no invented numbers—if a metric is missing or N/A, the model is told to say so.

What is sent to the model

build_metrics_prompt() serializes the metrics dict into a fixed Markdown-style block, including when present:

  • Duration, distance
  • Max total / horizontal / vertical speed
  • Takeoff and max altitude (MSL), max above takeoff, altitude gain
  • Max gravity-compensated acceleration
  • Energy (mAh), average current
  • Average satellites
  • Warning strings: distance, altitude gain, battery, gyro extremes

If ekf_available is true, EKF max speeds (total, horizontal, vertical) are appended.

The user custom mode prepends your question, then appends this metrics block. Other modes prepend a task-specific instruction (“summarise…”, “produce a full report…”) before the metrics block.

UI behaviour

  • Run: Click Analyse flight with Gemini (disabled in custom mode until the question is non-empty). A spinner runs while generate_content executes.
  • Errors: Exceptions are shown as Markdown starting with **API Error:**.
  • Cache: Results are stored in st.session_state under a key that includes file_key (which log is active), mode, and custom question text, so switching files or modes does not overwrite unrelated results.

Privacy and cost

Sending a report invokes Google’s generative AI API with the metrics text above. Treat the API key like any secret; restrict keys in Google Cloud / AI Studio as you would for production. Usage may be metered or billed according to your Google account and model pricing.