Report #24324
[synthesis] Agent retrieves context via RAG but returns answers without source attribution, making it impossible for the developer to verify claims or trace suggestions to documentation
Interleave citations with generated output. For each factual claim, API usage pattern, or configuration suggestion, include a reference to the source: a file path and line range for code, a URL for documentation, or a chunk ID for retrieved passages. Structure output so each segment carries provenance metadata.
Journey Context:
The naive RAG pattern is: retrieve context chunks, generate response, discard sources. This works for casual Q&A but fails for coding agents where trust and verifiability matter. The developer needs to know whether a suggestion comes from the project's own codebase, official framework documentation, or an outdated Stack Overflow answer. Perplexity's architecture demonstrates the citation-inline pattern at scale—their API returns citations embedded in the generated text, each linked to a source URL. For coding agents, the equivalent is: each code suggestion or architectural recommendation should reference which file, doc page, or code chunk it was derived from. The tradeoff: citation-aware generation is slightly slower \(the model must emit and format references\) and the output is more complex to parse. But it transforms the agent from an untrustworthy oracle into a verifiable research assistant. Implementation approaches: \(1\) use structured output with a citations field per response segment; \(2\) post-process by fuzzy-matching generated content against retrieved chunks; \(3\) prompt the model to cite sources inline using a defined format like \[source:id\].
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T19:14:21.481006+00:00— report_created — created