Manent is grounded in science

Every indexed excerpt carries its locator (file, section, page — down to the timecode of a video) from the moment of extraction, never reconstructed after the fact. Every answer displays its sources and opens the original at the cited page.

Grounding generation in retrieved documents (retrieval-augmented generation) reduces factual errors and makes every answer verifiable by a human.

Lewis, P. et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS.

Every answer is broken down into claims, and each one is verified against the cited sources. An unsupported claim is flagged — never papered over.

Language models produce plausible but sometimes unfounded statements — a risk documented by research, which calls for verification independent of the generation. A 2026 study (DELEGATE-52, Microsoft Research) shows that even the best models silently distort documents over long tasks — which makes this verification, on every answer, indispensable.

Ji, Z. et al. (2023). Survey of Hallucination in Natural Language Generation. ACM Computing Surveys.
Laban, P. et al. (2026). LLMs Corrupt Your Documents When You Delegate (DELEGATE-52). Microsoft Research.

Every answer is classified: extracted word for word from your documents, a synthesis marked as interpretation, clarifying questions, or an honest acknowledgment of a documentary blind spot. Safety content is never rephrased — it is cited exactly as written.

A reliable prediction system must be able to abstain when its confidence is insufficient (selective classification); models carry a confidence signal that can be exploited to this end. Conversely, a model that does not know when to stop drifts silently — this is the mechanism brought to light by DELEGATE-52.

El-Yaniv, R. & Wiener, Y. (2010). On the Foundations of Noise-free Selective Classification. Journal of Machine Learning Research.
Kadavath, S. et al. (2022). Language Models (Mostly) Know What They Know.
Laban, P. et al. (2026). LLMs Corrupt Your Documents When You Delegate (DELEGATE-52). Microsoft Research.

Every question queries both a lexical index and a semantic index, fused then re-ranked — the exact code "RR42" and "the machine makes a weird noise at startup" both find their page.

The BM25 probabilistic framework remains one of the most robust foundations of document retrieval; combining it with semantic search covers both families of field questions.

Robertson, S. & Zaragoza, H. (2009). The Probabilistic Relevance Framework: BM25 and Beyond. Foundations and Trends in Information Retrieval.

No knowledge card becomes official without the approval of YOUR experts. Every publication is reversible — a withdrawn card disappears from answers immediately.

Human-AI interaction guidelines recommend keeping human control over consequential actions and offering efficient, reversible corrections.

Amershi, S. et al. (2019). Guidelines for Human-AI Interaction. ACM CHI.

When the documentation has no answer, the conversation becomes a draft of a structured card (symptom, cause, fix) submitted for validation. The knowledge of your best people becomes writing — and writing stays.

Critical organizational knowledge is largely tacit: it lives in the heads of experts and is lost when they leave if it is not made explicit and structured.

Nonaka, I. & Takeuchi, H. (1995). The Knowledge-Creating Company. Oxford University Press.
DeLong, D. W. (2004). Lost Knowledge: Confronting the Threat of an Aging Workforce. Oxford University Press.

From the arrival inventory (each file, its status, its quote) to the audit report, every document, excerpt and answer is traceable. Nothing enters or leaves the system without a trace.

The credibility of a synthesis rests on the traceability of each element back to its source — the discipline that PRISMA 2020 imposes on scientific reviews, applied here to technical documentation.

Page, M. J. et al. (2021). The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ.

Three scenarios (conservative, typical, favorable), every assumption visible and adjustable, and signals measured on your corpus. After go-live, the actual return measured in use replaces the projection.

Structure inspired by the public principles of Forrester's Total Economic Impact™ methodology — risk-adjusted benefits, transparent assumptions (public text: forrester.com/policies/tei). Manent is neither affiliated with nor endorsed by Forrester.
McKinsey Global Institute (2012). The Social Economy. "Interaction workers" — managers, engineers, professionals and administrative staff — spend nearly 20% of their week searching for internal information or the person who knows; a searchable knowledge base can recover up to 35% of that, roughly 6% of the week returned to value-adding tasks.

An isolated space per client (your data never crosses another's), hierarchical roles, an event log, and both the search and the erasure of a person's information on request.

The Québec framework for the protection of personal information: consent, minimization, right to erasure, demonstrable accountability.

Loi 25 (LQ 2021, c. 25) — Act to modernize legislative provisions as regards the protection of personal information, Québec. Official text on LégisQuébec; all provisions have been in force since September 2024. General information — this is not legal advice.

The assistant follows the technician on mobile: question dictated, sourced answer, a card guided one question at a time. Augmented-reality glasses are being prepared — the phone remains the brain, the glasses a display.

In-task, hands-busy assistance is an axis documented since the origin of industrial augmented reality.

Caudell, T. P. & Mizell, D. W. (1992). Augmented Reality: An Application of Heads-Up Display Technology to Manual Manufacturing Processes. IEEE HICSS.