The context-strategy engine

The central decision of any RAG is which context travels to the model. karajan-rag does not impose it: over the same index it offers three strategies (--mode rag|cag|hybrid), all through the same guarded path (sensitivity policy + PII redaction). rag is the default you have already seen: the top-k chunks from hybrid retrieval.

CAG mode: the full corpus as context

For small/medium corpora, --mode cag (Cache-Augmented Generation) skips retrieval and loads the whole corpus into the model context:

karajan-rag query "summarize the architecture" ./my-project --answer --mode cag

• The context is deterministic and stable (ordered by path): same corpus → same prompt, which lets the provider amortize its prompt cache across queries.
• The effective sensitivity is the maximum across the whole manifest — the entire corpus travels here, so the gate is stricter than in RAG by design. PII redaction applies just the same.
• Budget with explicit failure (--max-context-chars, default 400K characters ≈ 100K tokens): if the corpus does not fit, an error reports the real size and the alternatives — never a silent truncation.
• It needs no vector store to answer (the manifest suffices), but it does need an indexed corpus.

And the middle ground, --mode hybrid: retrieval selects the relevant files and the context carries those files whole (not fragments), with the ones that do not fit the budget declared in the log. Ideal when chunks fall short but the whole corpus does not fit.

Rule of thumb: corpus that fits in context and questions asking for a global view → cag; large corpus and pinpoint questions → rag; large corpus and questions that require understanding entire files → hybrid. Unsure about YOUR corpus? Decide with data:

karajan-rag eval golden.json --compare-modes

It compares offline the retrieval recall against the context cost of each mode and emits a recommendation justified with numbers.