InsightForge is an educational / portfolio project (an Applied GenAI capstone). It runs over a synthetic sales dataset and public BI literature — there are no real customers, credentials, or secrets in this repository.

If you find a security issue you'd still like to report, please use GitHub's private vulnerability reporting (the repository's Security tab → Report a vulnerability) rather than opening a public issue. I'll respond as time allows — this is a personal project, not a maintained service.

• The LLM never invents a number. Every statistic is computed deterministically by pandas (analytics.py); a custom retriever feeds the relevant figures to the LLM, and a post-generation numeric-grounding gate (grounding.py) rejects any answer whose figures don't trace to the retrieved statistics — falling back to the deterministic rendering. Backed by tests.
• Citations are attached in code, not authored by the LLM. Qualitative recommendations are grounded in retrieved passages from the source PDFs, cited with source + page from the retriever's metadata (rag.py) — the model can't fabricate a source.
• No secrets committed. .env is gitignored; the Anthropic API key lives only locally. The app runs fully offline (USE_LLM=0) with no key.
• Read-only data path. The CSV is loaded read-only; there is no write path, no user accounts, and no PII beyond synthetic age/gender columns.

Documented so they aren't mistaken for oversights; a production deployment would close each one:

• Retrieval is TF-IDF, not semantic embeddings. Deterministic, offline, and CI-friendly (no torch), which suits a demo and keeps tests reproducible. Production would use a semantic embedding model + a managed vector store for better recall on paraphrased questions.
• Single CSV, in-process. Fine for a demo; production would use a warehouse/connector and per-tenant access controls.
• Memory is an in-process conversation buffer. No cross-session persistence or summarization (documented in the README).
• The numeric gate is a high-recall heuristic. It catches fabricated or LLM-computed figures by value-matching with a small rounding tolerance; it is not a formal proof of correctness (the underlying figures are correct by construction from pandas). It checks a figure's provenance, not its binding to the right label — correct attribution is the LLM's job, constrained by the exact-quote prompt.
• Numeric grounding is numeric-only; the qualitative source corpus is trusted. The gate doesn't filter prose, and the BI source PDFs are trusted operator input (no upload path — SOURCES_DIR is fixed). An adversarial PDF could inject a qualitative instruction; out of scope for a curated single-operator demo.