grounded-generation

Large-scale pretrained models for goal-directed dialog

Open-vocabulary Object Segmentation with Diffusion Models

This is a list of open-source projects at Microsoft Research NLP Group

Statewave use case in TypeScript: a grounded product advisor and support assistant, with an Ops Assistant that tracks and closes content gaps.

Deterministic multi-LLM orchestration with grounded, verified answers, an editable guardrail gate, and a live process dashboard.

Financial report grounded QA post-training with QLoRA SFT, DPO preference learning, citation/numeric evaluation, and badcase analysis.

面向建材销售与技术支持的本地多模态 RAG Agent：异构文档解析、混合检索、原图证据回传、LangGraph 工具规划与 16GB GPU 部署

Beginner-friendly Python RAG project that turns StackOverflow Q&A data into a grounded knowledge assistant using sentence-transformer embeddings, FAISS retrieval, and strict citation-based LLM answering.

HEFEST — on-prem рабочее место по химической безопасности завода: паспорта ГОСТ 30333, аварийные карточки, зоны АХОВ, совместимость хранения, СИЗ, класс отхода. Поиск — фундамент, а не продукт. Офлайн, с провенансом, без выдуманной нормативки.

Conversational + predictive crime intelligence on Zoho Catalyst: ask 1M+ FIRs in English or Kannada and see the query and rows behind every answer. All-India NCRB-calibrated corpus (640 districts), forecasting validated on 6.37M real incidents, and a trilingual WhatsApp channel for field officers.

Levi's RAG is an AI due diligence copilot for equity research on Levi Strauss & Co. It answers natural-language financial questions using only official SEC filings, grounding each answer in the exact source sentence and tagging it with an evidence tier (verified, qualitative, benchmark, or inference).

ComfyUI Custom Nodes for Gen Searcher and other tools

An open-source LLM for food product development

Grounded, evidence-cited crime-intelligence assistant over synthetic CCTNS-aligned records — every fact cites the FIR behind it, and it abstains when the records do not support an answer.

Developer-first prompt engineering patterns for grounded, testable, and reliable AI outputs.

Secure Enterprise RAG platform with RBAC enforcement, OWASP LLM defenses, hybrid retrieval, grounded generation, explainability, and zero-trust context pipelines.

Dependency-free checks for structured local-model output, grounded answers, citations, and protected writes.

pip install cairn-assistant · Grounded-or-silent reference assistant for public agencies: answers only from a supplied corpus, cites every claim, and refuses cleanly rather than guessing. No model and no network at answer time. English and Spanish plus right-to-left languages. A reference implementation to read, not a product to deploy.

Strictly-grounded RAG chat on Adobe Spectrum docs — a native react-instantsearch (useChat) frontend on Algolia Agent Studio + an Algolia index. Per-answer grounding judge (3-judge corroboration-gate panel) as a Confidence chip → breakdown drawer. Reusable Algolia-Central-[company] template.

Turn research PDFs into grounded Markdown drafts — driven by any token-free AI coding-agent session (Claude, Codex, Gemini, Copilot, Cursor, Amp, Crush, Goose, Grok) when online, local Ollama when offline.