From raw feedback to a cited PRD
A retrieval-augmented pipeline where every generated recommendation is grounded in real user quotes — the citation trail is the product.
Paste feedback or upload a CSV. The parser handles a text column plus optional source/date.
Each item becomes a vector. Mock mode uses deterministic TF-hashing embeddings; live mode uses OpenAI text-embedding-3-small.
Greedy cosine-similarity clustering groups feedback into themes — deterministic, so the same upload yields the same themes.
Each theme gets an opportunity score from size + negative-sentiment intensity.
For a question or a PRD, cosine search fetches the most relevant feedback as cited evidence (the R in RAG).
PRDs and answers are generated grounded in the retrieved evidence (claude-opus-4-8 in live mode), with a ≥2-citation guardrail.
Anthropic has no first-party embeddings API, so live mode uses OpenAI for vectors and Claude (claude-opus-4-8) for generation. Mock mode removes both — deterministic hashing vectors mean the demo runs with zero keys.
Why citations are a guardrailAn evidence-first tool lives or dies on trust. Every PRD must cite at least two sources; the generator refuses otherwise. The metrics view tracks the share of feedback that couldn't be grounded into a supported theme.