Decision in brief
Use SearXNG when you want web results from multiple upstream engines through an instance you operate. Build a private index when you need to search your own records or cannot send queries to outside services. Keep conventional browser search if upstream failures in the pilot make the extra service hard to justify.

Separate web results from documents you own
SearXNG sends web queries to configured search engines and combines their responses. Your deployment controls local access and logging. Upstream engines still receive relevant queries from the instance and can time out, throttle or return no results. Review settings for outgoing timeouts and enabled engines, and keep sensitive internal searches on a different endpoint.
A private index serves documents you feed into it. Use a file or history search tool for one person's approved local data. For a team catalog, consider a managed index or a self-hosted service such as Meilisearch and plan ingestion, deletion and permission evaluation. Keep admin credentials out of any browser integration.
| Search task | Starting point | What to measure |
|---|---|---|
| Public web with self-managed interface | SearXNG with a small engine allowlist | Top-five relevance, empty responses and latency by query group. |
| Offline private files | Local index over a named folder or corpus | Expected results without external connectivity and removal after deleting a source. |
| Team documents with permissions | Access-controlled indexing with a filtered query path | Every permitted result appears and prohibited documents stay hidden for the test identity. |
| Standard personal web search | Existing browser default as a baseline | Switch only if relevance and reliability make the new operational burden worthwhile. |
Two HN cases showing different failure modes
In an AgentSearch thread, tahseen0amin reported that a SearXNG-backed service responded but returned zero results. Include a known-answer smoke query to distinguish a running interface from a functioning upstream search path.
dandaka described building Traul to index messages from several apps in a local database. In that design retrieval quality depends on the set of ingested sources and the freshness of the sync. Review what enters the index and rehearse removing a test result from it.
Evaluate a week of representative queries
- Prepare 20 harmless searches with an expected result or relevance rubric. Keep public queries and synthetic private documents in distinct sets. If the use case is web search, compare your browser default and private metasearch instance.
- Record top-five hits, latency, empty responses and upstream timeouts under ordinary use. Retest with one configured engine intentionally disabled. If using a document index, test insert, update, deletion and two roles.
- Verify browser use from the intended devices, review local request logs for query retention, and restore one configuration or rebuild the synthetic index from source.
- Proceed if at least 18 of 20 suitable queries find expected results in the top five, the chosen latency target holds and every private visibility test passes. Stop if private search requires web requests or any forbidden document appears.