Ollama Web Search

Original Article ↗ Hacker News Discussion ↗

Perceived benefits of Ollama Web Search

  • Many see web search as solving a key weakness of small/local models: lack of up‑to‑date or niche knowledge.
  • Users report surprisingly good results, including for “deep research” when paired with larger models.
  • Some like being able to cheaply test large models in the cloud before deciding whether to run them locally.

Comparison to local search / MCP alternatives

  • Several people already use SearXNG, Tavily, SERP API, or DuckDuckGo/Google Programmable Search wired into their own agents.
  • SearXNG with Open WebUI and large open models is described as “good enough,” though sometimes slow; others say tweaking timeouts and engines helps.
  • Some argue a local SearXNG + local LLM stack removes the need for Ollama’s hosted search.

Cloud offering, pricing, and business model

  • Confusion and pushback that a tool marketed for local models now requires accounts and sells hosted models/search.
  • Debate on why pay for large open models on Ollama Cloud instead of frontier proprietary models from major providers.
  • Counterargument: open models are rapidly improving, cheaper, and customizable; $20/month for access to several huge open models via a local-compatible API is seen as good value by some.
  • Skepticism about sustainability of flat‑rate pricing and concerns about inevitable “enshittification” under VC pressure.

Search backend, licensing, and privacy

  • Strong interest in which search providers are used (Brave, Exa, etc.) because their ToS often restrict storing or republishing results.
  • Ollama says results are “yours to use” with zero data retention, but refuses to name providers or detail licensing; this is seen as legally and practically unclear.
  • Lack of a clear privacy policy at launch and CCPA implications are flagged as red flags.

Local vs hosted search implementation

  • Ollama says they tried fully local crawling but hit quality issues and IP blocks; hosted APIs were a faster path. They say they still “believe in local” and may revisit.
  • Some users see the account requirement as “dead on arrival” and are migrating to alternatives like llama.cpp, vLLM, or RamaLama, especially for on‑prem use.

Tool use, integration, and enterprise search

  • Web search works with tool-capable local models via their tool API; examples include using qwen or gpt‑oss and wiring search as an agent tool.
  • Some want Ollama to prioritize robust local tool use instead of hosted search; Ollama claims tool support has been improved.
  • For enterprise/local search, suggestions include Solr (with MCP integration and vector search), Typesense, and Docling; others run hybrid systems (LibreChat + llama.cpp + Tavily, etc.).

Broader search & ecosystem debates

  • Discussion branches into the economics of web indexes, feasibility of P2P or mini‑Google setups, and how AI‑mediated search might threaten ad‑driven search engines.
  • Several note that Ollama is one of many interchangeable components now; if it drifts away from its local/OSS positioning, users can and will swap in other backends.