Watch filters: what they are, how they work and why they decide the quality of a watch

What exactly is a watch filter?

In information retrieval, a standing query is one that is run periodically over a collection to which documents keep arriving. The classic example in the textbook by Manning, Raghavan and Schütze is a query about ‘multicore chips’ run every morning against wire news: if it is not expanded with ‘multicore processors’, it loses relevant articles^[1]. The same text describes it as a classification based on rules written by hand, often equivalent to Boolean expressions, such as (multicore OR multi-core) AND (chip OR processor OR microprocessor)^[1].

Its great advantage is transparency: anyone can read a filter and understand why one publication came in and another did not. Its cost is maintenance: a sector’s vocabulary changes, and a filter that is not reviewed degrades silently.

Why can precision and recall not be maximised at the same time?

Precision is the fraction of what is retrieved that is relevant; recall is the fraction of what is relevant that is retrieved. The two trade off: returning everything gives perfect recall with negligible precision, and recall never drops when a query is broadened, whereas precision usually does^[2].

In watching, the balance depends on what each watch is for. Regulatory watching for a product subject to CE marking tolerates more noise in exchange for not missing any standard, because a false negative can cost a non-conformity. A trends watch for the executive committee needs the opposite: few signals and highly relevant ones, because the reader stops opening a saturated newsletter.

Which operators are used, and what effect does each one have?

Operators are common to most professional databases, although the syntax varies between tools. Espacenet, the patent database of the European Patent Office, for example, offers operators to require all terms, any of them, an exact phrase or proximity between terms; it does not allow AND, OR and NOT within a single search field, only between fields^[3].

How to build a filter that stands the test of time

1. Write the need first. Which decision the watch must support, and which publication would be a signal: ‘any regulation or draft regulation that restricts the use of PFAS in food packaging’.
2. Separate the core concepts. In the example, three: the substance, the application and the regulatory character. Each becomes a block, joined to the others with AND.
3. Complete each block with variants joined by OR: synonyms, acronyms, trade names, plurals and translations, if you watch sources in other languages.
4. Use exact phrases for compound terms that lose their meaning when separated.
5. In patents, combine words and classification. The Cooperative Patent Classification (CPC), managed by the EPO and developed with the USPTO, groups documents by technical subject regardless of the words each applicant uses. Its version 2026.08 entered into force on 1 August 2026^[4].
6. Exclude at the end, and with care. A NOT is only added when the same noise keeps recurring, and the reason is noted.

How to tell whether a filter works

• Test it with known cases. Gather publications that the filter should catch and others it should leave out, and check before activating it.
• Note the reasons for discarding during the first weeks. If they repeat, the adjustment lies in the filter.
• Look for false negatives. What the filter leaves out is not visible: when a relevant signal arrives by another route (a customer, a congress, a colleague), check why it did not come in.
• Review the vocabulary from time to time. New acronyms, new players and new trade names appear without warning.

Filters and AI at Innguma: perimeter and order

At Innguma, each watch’s filters define the perimeter: what comes in. Within that perimeter, Innguma’s AI filters, groups duplicates and classifies each signal by category, technology, entity, type of event, place and tone. Rules give explicit control; classification helps to sort when the same technology is named in many ways. People decide what is watched, what is published and who receives it.

If your organisation enables it in its integration, the filters of the watches can also be modified from a compatible assistant through the Model Context Protocol (MCP): ChatGPT, Claude, Microsoft Copilot or Gemini. Which changes are automated, and which go through review first, is decided by your team. The configuration details of each filter are defined with each organisation during set-up.