Launch HN: Danswer (YC W24) – Open-source AI search and chat over private data

Hacker News Discussion ↗

An open-source project called Danswer aims to provide AI-powered search and chat over a company’s private knowledge, aggregating content from tools like Slack, Google Drive, Confluence, GitHub and more. Commenters probe how it handles retrieval quality, access control, multilingual support, and incremental syncing at scale, and contrast it with both in-house RAG prototypes and commercial offerings such as Microsoft Copilot or Glean. The founders position openness, self-hosting, extensible connectors, and a focus on retrieval quality as their main edge, while monetizing through a paid cloud version and advanced enterprise features.

Product Focus and Differentiation

  • Positioned as an open-source, team-focused “unified search + AI chat” over private company data, not just a ChatGPT-like UI.
  • Emphasis on the retrieval half of RAG: connectors to tools like Notion, Confluence, Jira, Slack, Google Drive, GitHub/GitLab, etc., incremental updates, metadata, and access control.
  • Compared to alternatives (OpenWebUI, Glean, dust, Vectara, privateGPT, Copilot): Danswer stresses open source, self-hosting, strong retrieval, and multi-source aggregation; others are seen as more proprietary, narrower, or more “assistant/agent”-oriented.

Open Source, Business Model, and Moat

  • Core is MIT-licensed; paid cloud and some advanced enterprise features (e.g., SAML/OIDC, expert-finding, advanced RBAC) are proprietary.
  • Strategy: win small and mid-size teams who are ignored by high-touch SaaS vendors, and appeal to large orgs that want transparency, customization, and self-hosting.
  • Competition from FAANG/Copilot acknowledged, but founders believe OSS + community + flexibility is a defensible angle.

Architecture, Models, and Retrieval Quality

  • Uses Vespa as vector DB + search engine; hybrid search (keyword + embeddings), local ~100M-parameter models on CPU for embeddings/rerankers, default OpenAI LLM but pluggable/open-weight models (including via Ollama).
  • RAG pipeline includes context-aware chunking (via limited LlamaIndex use), multi-granularity passes, feedback-based re-ranking, and time-based decay.
  • Novel reranking approach: use an LLM to judge whether chunks are “useful,” not just “relevant,” claiming better results than standard cross-encoders.

Security, Privacy, and Access Control

  • Multiple auth options (basic, Google OAuth; OIDC/SAML in paid tier). Domain-restricted signups supported.
  • RBAC currently connector-level; ongoing work to sync fine-grained permissions from sources (Google Drive, Confluence, Jira, Notion, Slack channel membership).
  • For SaaS, admins may access data under agreement; for self-hosted, infra admins typically already have access to source systems. Air-gapped setups with local LLMs are possible.

Limitations, Roadmap, and Use Cases

  • Current gaps: no full code search yet (only PRs/issues), basic spreadsheet and PDF handling, no OCR or rich table/graph understanding; these are on the roadmap.
  • Real-time/streaming ingestion is partly scaffolded but mostly unused due to source API limits.
  • Strong interest and positive feedback from teams with painful internal search, especially consulting and enterprise/government; some skepticism about LLM reasoning depth, UX around private channels, and long-context knowledge exploration.