Analysis Modes

Oversight supports three analysis modes that determine which LLM providers are used for hallucination, bias, and toxicity detection. You can set a default mode in Settings and override it per upload.

Gemini Mode

Provider: Google Gemini (gemini-2.5-flash by default)

• Balanced accuracy and speed
• Best general-purpose option for most use cases
• Only requires GEMINI_API_KEY
• Includes automatic model fallback: if the primary model is rate-limited, falls back to gemini-3.1-flash-lite

When to use: Default choice for most analyses. Good accuracy with reasonable speed.

Groq Mode

Provider: Groq (Llama models, openai/gpt-oss-120b by default)

• Faster inference than Gemini
• Requires GROQ_API_KEY
• Also used for the live chatbot and live monitoring
• Includes automatic model fallback chain: if the primary model is rate-limited, falls back through Kimi K2 variants and compound models

When to use: When speed is a priority, or when you want to compare results against Gemini analysis.

Both Mode

Providers: Gemini first, then Groq cross-check

• Highest confidence results
• Gemini analyzes the conversation first
• Groq then performs an independent cross-check
• The cross-check prompt includes Gemini's findings for validation
• Issues confirmed by both models receive the highest confidence scores
• Requires both GEMINI_API_KEY and GROQ_API_KEY

When to use: When accuracy is critical and you want maximum confidence in the results. Takes longer since both providers run sequentially.

Comparison

Feature	Gemini	Groq	Both
Speed	Medium	Fast	Slower
Accuracy	Good	Good	Best
API keys needed	GEMINI_API_KEY	GROQ_API_KEY	Both
Cross-checking	No	No	Yes
Model fallback	Yes	Yes	Yes

Setting Your Default Mode

1. Go to /settings
2. Under Default Analysis Mode, select Gemini, Groq, or Both
3. Click Save

Your default mode is used for all new uploads unless overridden on the upload page.

Overriding Per Upload

On the /upload page, you can select a different mode for a specific upload without changing your default setting. This is useful for one-off comparisons or when you want to cross-check a specific conversation.

Model Fallback Chains

Both Gemini and Groq have built-in fallback chains to handle rate limiting gracefully:

Gemini fallback:

1. gemini-2.5-flash (primary)
2. gemini-3.1-flash-lite (fallback)

Groq fallback:

1. openai/gpt-oss-120b (primary)
2. Kimi K2 variants (fallback)
3. Compound models (final fallback)

If all models in the chain are rate-limited, the analysis for that category fails gracefully — other categories may still succeed. An upload is marked "completed" as long as at least one analysis category succeeds.