Hello Code4Lib,
I’m Michael Shamanoff, founder of Achla AI. I’m working on a managed
citation-first search layer for public library, archive, museum, and
research websites. This is a commercial project, which I want to disclose
up front.
The design constraint is narrow: an answer should only be presented as
answered when it is supported by identifiable pages or documents from the
site. When the corpus is insufficient, the interface should show an
explicit miss rather than a plausible unsupported answer.
Before proposing any pilots, I would value practitioner criticism of the
evaluation design:
- What evidence standard should a cited answer meet?
- How should conflicting or outdated institutional pages be exposed?
- Which PDF, digitized-object, or metadata formats most often break
retrieval?
- For multilingual questions, which terminology failures matter most?
- When does answer-plus-sources add value beyond conventional site search?
I wrote a short technical note describing the proposed evaluation framework:
https://researchsearchnotes.hcommons.org/2026/07/30/from-search-results-to-cited-answers-a-small-experiment-for-research-websites/
Pointers to existing work, prior failures, or reasons this approach may not
fit library systems would be especially useful. I will not collect
respondents’ addresses for other outreach without their consent.
Disclosure: I built the commercial prototype discussed in the note, and the
note was drafted with AI assistance and substantially reviewed and edited
by me.
Best,
Michael Shamanoff
Founder, Achla AI
https://achlaai.com
Received on Thu Jul 30 2026 - 00:59:21 EDT