What do people use for search?
I am trying to solve a pretty (in my head) simple use case. Intake a list of companies, proceed to make search queries about these companies (news, announcements, results) for articles posted within the past 7 days and dump title, snippet, url, etc into a…
I am trying to solve a pretty (in my head) simple use case. Intake a list of companies, proceed to make search queries about these companies (news, announcements, results) for articles posted within the past 7 days and dump title, snippet, url, etc into a file for later processing. Silly me, apparently search is really really hard even in 2026. So far I've tried: Exa, Tavily, Serper, Serpbase, Firecrawl, SearXNG and some others and none seem to produce anything even remotely acceptable. This is a big one, vast majority of search backends either outright do not support "freshness" or produce bad to non-existent results when you try to employ it. Meanwhile I can go to Google, make the same exact query and get the desired results. With Google I can enter "COMPANYNAME news announcements results" as a single query and get decent results. With various search backends, I seem pigeonholed into making 3 separate queries to get anything even remotely reasonable. Is this a deliberate tactic to get people to burn through their API credits? Results are often cached? With self-hosted models, I feel like I went 2 years back in time and this is acceptable to me. With search, however, I feel as if the jump is 30 years back, something of the Altavista age. How is any of this acceptable? How are people PAYING MONEY for this quality? What are the big boys using for their searches, Google deals behind closed doors (Google no longer offers search API directly)? What are you using and how did you have to wrangle with it to get acceptable behavior of it?
已收录讨论
searxng is free
it's also mentioned in the OP
Brave will detect automated browsing in a heartbeat. If you have something that can defeat a captcha it might work.
Does this do something not already done by web-search-plus?
Searxng uses google...how does it suck too?
There are many tools that do similar things and there is for sure overlap. This one is mine. The main difference is I wanted a tool that runs as a stateless CLI tool, and is kind of a Swiss Army knife of search and scrape tools. Ketch also is Web Search, Code Search (GitHub), and docs search under one roof. Plus it includes scrape tools, a crawl tool, and built in browser for those hard to scrape JavaScript sites. Try it out, let me know what you think. If there’s an area in which I can build something to improve your workflow I’m open to suggestions.
Yes, an llm powered research loop with grounded date/time. That's what I do too. Searxng is just one part of it.
I think this is the problem for why he was not getting the last week news correctly.
Hi. I no longer remember the finer details, but it was either that Exa did not respect "freshness" (which could also have been a problem with web-search-plus that sits as the intermediary and should know how to make the appropriate API call) or the results of queries for Finnish language news about slightly more obscure Finnish public companies were unsatisfactory.
Thanks, that was enough to narrow it down! I looked into web-search-plus by robbyczgw-cla, and its freshness="week" option was not actually being translated into Exa’s publication-date filters :/ (I am going to fix this with a PR to their repo) I recommend testing Exa directly on recent Finnish company news using explicit startPublishedDate and endPublishedDate values, I tried this myself and it worked great eg: { "query": "COMPANY NAME TICKER uutiset tiedotteet tulos", "type": "auto", "numResults": 10, "startPublishedDate": "2026-07-16T00:00:00Z", "endPublishedDate": "2026-07-25T00:00:00Z", "contents": { "highlights": true, "maxAgeHours": 24 } } The date fields control which results are selected. maxAgeHours controls how recently the selected pages are crawled Hit me up if you need some Exa credits to test us out properly :)
I ended up building ketch for this exact frustration. It gives me one CLI/MCP interface across Brave, Exa, SearXNG, etc., so I can swap or rotate backends without rewriting the research workflow. It doesn’t magically make a weak provider’s freshness better, but it makes comparing them easy. In the latest release I also added a —multi flag that lets you search across all your configured backends and return ranked results. I’m also very active on Reddit. Issues and PRs welcome if you want to add backends or features! https://github.com/1broseidon/ketch
Hey! I work at Exa. Would love to understand where we fell down here. Would you mind sharing one or two example companies plus the exact query/API call you made that produced poor results? If you still have the JSON response, even better :) The workflow you're describing (recent company news) is something we actively care about and a lot of customers use us for it, so I'd love to figure out whether it's an issue on our side or something else.
I use openrouter's search plugin. You can also use talordata, serper.dev, i think even firecrawl can do searches.
I usually get good search results using 4 things: SearXNG as the search provider Enable a time/date tool A model with good tool call, the one that gave me the best results was Qwen3.6 35B Max reasoning So what happens, if I ask the model to search for something that happened yesterday, it uses the time_date tool to get the date, makes searches using the date, and if it doesn't find the answer, it does another search or fetches a page for more info. I tried Gemma 4, and they were bad with tool calls, often they don't search the web until I tell them too, and they don't use the time_date tool first, leading to old data. Qwen 3.6 seems to get that logic right. You can test the models with "Who won the X sport game between A & B yesterday", or how did company X and Y stock price close yesterday. If they don't work well, you have to be more explicit, like use the get_timedate first to get the current date, then use the web search and fetch tools to search for news in the last day/week...
I have a simple duck duck go tool. Maybe doesn’t suffice though
You can enable tool calling in llama.cpp and it should be able to search. You can also add brave api
SearXNG + paid Brave search api
Brave search api has my vote
I don't know man searxng results are superb https://pastebin.com/P5AXP5t1 that's requesting a comparison across 3 $90 thermal cameras on amazon and the report is solid
Tavily free tier here.
This is where local models fall off for me. They don't do a good job of pulling in up to date information due to the limitations of searching/scraping the web with currently available tools.
I have this same question, trying to get small abliterated models to run research search (on a 10gb RTX 3080), and haven’t gotten anything to work well.
I use Brave search it’s free
Shameless plug, but I built a OS tool for part of this problem https://github.com/silentrob/webrag After you have a URL you want to get context for, this fetches the page and removes the useless extra tokens.