Sift Dashboard Settings Help No password set
Setup needed. The search API key is missing. The AI API key is missing. Runs cannot collect anything until both are set — add them in Settings.

How Sift works

A short guide to the whole tool. Nothing here requires technical knowledge.

The idea in one paragraph

You tell Sift what to search for and what information you want. It searches Indian news, opens each article it finds, and pulls out the details into a spreadsheet. Every row links back to the article it came from, so you can always check it.

Getting set up (once)

  1. Add your two API keys in Settings. One for searching (ValueSERP), one for the AI that reads articles (OpenRouter). Nothing can be collected until both are in.
  2. Set a sign-in password in Settings, before this is reachable from the internet.
  3. Update the model list in Settings so Sift can tell you what any AI model costs.

Running a collection

  1. Choose a template on the Dashboard — it decides the spreadsheet columns. Edit the columns first if you want something different.
  2. Press Start a run and fill in the four steps: what to search for, which regions, how much to collect, and which AI model.
  3. Read the cost estimate at the bottom. It updates as you change things, so you can see what a bigger target or a different model would cost.
  4. Create the run. It joins the queue and the background worker starts it within a few seconds. Reload the run page to watch progress.
  5. Download the spreadsheet from the run page when it finishes — one file per year, as Excel or CSV.

Things worth knowing

It is safe to stop Turning the worker or server off mid-run loses nothing. When it comes back it carries on from where it stopped, without redoing or skipping work.
Repeating a run is nearly free Searches are remembered, so running the same thing again costs almost nothing.
Duplicates are merged When several outlets report one event you get a single row listing all of them, with a higher confidence score.
Confidence Higher when more outlets reported the same event and more fields were filled. Single-source rows are kept, just marked lower.
The review list Anything that fails a check goes there instead of into the spreadsheet, so nothing is quietly dropped or quietly wrong.
Older news is patchier Many articles from 2013–2015 are no longer online, so early years return less than recent ones. The run page shows which regions and years came up short.
Sources are news sites only Search and collection are restricted to a news-domain list. Social media and government sites are never fetched, whatever the search terms are.

If something looks wrong

The run stays "queued" The background worker is not running — it is the part that actually collects. Close this window and start the program again with start_program.bat, which starts both halves. Nothing is lost: the run picks up where it stopped.
Very few records Usually too few search words or too narrow a date range. The run page shows how many articles were found, downloaded and accepted at each step.
A field will not save Two columns share a name, a name is one the platform already supplies, or a judgement column has no quote column. The message names the field and the reason.
Cost looks wrong If you typed a custom model name, update the model list in Settings so it can be priced. Until then the estimate leaves out the AI cost and says so.