How are the diffs being analysed?

[pantharshit007]

Git provider (optional)

Github

System Info (optional)

No response

Issues details

Could you provide details on how diff analysis is handled in the system?

Specifically, is there a tiered approach where more cost-efficient models are used for initial or lower-priority analysis, and more advanced models are employed for complex or higher-level reasoning tasks?

Using self-hosted PR-AGENT
Models: Open Router and Open source models via Ollama (for a cheaper alternative)

[qodo-merge-for-open-source]

💡 Auto-generated documentation-based answer:

How diffs are analyzed:

• Dynamic/asymmetric context: Qodo Merge expands the unified diff with more lines before than after and, when possible, includes the enclosing function/class (dynamic context). This helps the LLM reason on the change in its true scope without flooding the prompt.
• Full-file retrieval when needed: The system can fetch entire modified files and use them to enrich prompts, plus AI-generated PR metadata (type, summary, per‑file “changes walkthrough”) from /describe is reused by /review and /improve to provide multi‑stage reasoning.
• Chunking for large PRs: For big PRs, the code is split into chunks that fit the model’s context, and suggestions are generated per chunk (configurable).
• Validation and self‑reflection: Suggestions are validated (location/syntax) and re‑ranked via self‑reflection to reduce noise and prioritize important findings.

Tiered models / cost strategy:

• Open source (PR‑Agent): You can explicitly set a primary model and fallbacks via config (config.model, config.fallback_models). There isn’t a built‑in “tiered pipeline,” but you can approximate one by:
  • Choosing a cost‑efficient model as primary and a stronger one as fallback.
  • Using OpenAI Flex Processing for background tasks to reduce costs.
• Qodo Merge (hosted): Defaults to a combination of top models (e.g., GPT‑5, Haiku‑4.5, Gemini 2.5 Pro). You can also pin a single model if desired. The docs don’t describe an automatic multi‑tier cascade beyond configurable primary/fallback selection.

What to provide if you need a more tailored answer:

• Which deployment you use (self‑hosted PR‑Agent vs Qodo Merge).
• Your model/provider preferences and cost constraints (so we can suggest a concrete primary/fallback setup).
• Which tools you care about most (/describe, /review, /improve), since each can have its own config.

Relevant Sources:

• https://qodo-merge-docs.qodo.ai//core-abilities/dynamic_context#supported-git-platforms:-github-gitlab-bitbucket
• https://qodo-merge-docs.qodo.ai//core-abilities/metadata#local-and-global-metadata-injection-with-multi-stage-analysis
• https://qodo-merge-docs.qodo.ai//tools/improve#how-many-code-suggestions-are-generated
• https://qodo-merge-docs.qodo.ai//usage-guide/changing_a_model#changing-a-model-in-pr-agent
• https://qodo-merge-docs.qodo.ai//usage-guide/changing_a_model#openai-flex-processing
• https://qodo-merge-docs.qodo.ai//usage-guide/qodo_merge_models#the-default-models-used-by-qodo-merge-💎-october-2025-are-a-combination-of-gpt-5-haiku-45-and-gemini-25-pro
• https://qodo-merge-docs.qodo.ai//core-abilities/self_reflection#self-reflection
• https://qodo-merge-docs.qodo.ai//core-abilities/code_validation#code-validation