Local Knowledge Cache

The Local Knowledge Cache (LKC) explores how place-based knowledge systems can prioritize traceability, social verification, and interactional context over fluent synthesis.

Rather than producing generalized answers about a place, the project investigates how knowledge can remain grounded in local usage, negotiated authority, and situated meaning, especially when accessed by outsiders.

LKC serves as the retrieval and reference layer for downstream design interventions within the Local Conversation Studio program.

LKC functions as the knowledge infrastructure of the program.

Its role is to translate qualitative evidence into structured reference material that can be consulted responsibly, without replacing conversation or judgment. The system is explicitly designed to support interaction and orientation rather than act as a local "answer engine."

Situated between analytic workflows and embodied interventions, LKC preserves provenance, context, and uncertainty while keeping local authority socially verifiable.

Visitors to Oaxaca often arrive without essential local context, including:

• etiquette and interaction norms
• language and vocabulary for food, greetings, and everyday conversation
• temporal knowledge (hours, festivals, civic rhythms)
• navigation norms and service access
• cultural taboos and usage constraints

As more visitors turn to generative AI systems such as ChatGPT for travel guidance, scholars have begun to note that generative travel advice often reproduces mainstream destination narratives and can overlook localized or emerging places. Studies of AI-generated tourism recommendations show tendencies toward simplified spatial representations and bias toward widely recognized landmarks, raising concerns that place-specific knowledge may become generalized or distorted (Andreev et al., 2025; Dudy et al., 2025).

Major policy guidance warns that generative AI systems trained on large-scale web data can reproduce dominant cultural norms while appearing authoritative, posing particular risks to linguistic and cultural diversity in underrepresented contexts (UNESCO, 2023).

This is not merely an accuracy problem, but an authority problem. Local knowledge is negotiated through interaction and social verification. Systems that generate answers without that grounding risk displacing local expertise rather than supporting it.

LKC takes a retrieval-first approach as a normative design choice, not simply a technical preference.

Instead of synthesizing answers, the system retrieves discrete, human-curated knowledge items with contextual metadata intact. Each item includes:

• provenance (where did this come from?)
• verification status (how widely is this supported?)
• appropriateness constraints (when and by whom does this apply?)
• acknowledged variants (are there competing interpretations?)

This approach allows social verification to continue shaping knowledge over time, rather than allowing fluent synthesis to quietly overwrite local context.

LKC was developed as a staged system, beginning with representational infrastructure rather than community-scale deployment.

Claims as the Atomic Unit

The fundamental unit of the LKC is a claim: a bounded piece of knowledge paired with metadata that supports verification and appropriate use. Claims carry information about provenance, social context, and variation, allowing users to inspect knowledge rather than consume it as fact.

Phase 1: Initialization with Web Sources

The initial phase focused on building schema and indexing capacity.

• Source acquisition: place-related material collected from diverse web sources
• Structured parsing: extraction of entities, key claims, terminology variants, and timestamps
• Evidence indexing: storage of hundreds of documents with preserved citations

This phase prioritized representational capacity over authority or completeness.

Phase 2: Participant Review & Social Verification (Pending)

The schema supports participant review, confirmation, and contestation, but these workflows have not yet been fully activated. Introducing local review is a central focus of future work.

Evaluation emphasized situated use rather than interface performance.

Situated Evaluation via Cards

When knowledge items were rendered as Local Language Cards and used in live interaction, participants corrected wording, flagged inappropriate usage, and suggested contextual refinements.

This revealed that use itself can function as a validation mechanism, surfacing errors that would not appear through abstract review alone.

1. Knowledge for interaction, not observation

47% of the knowledge store's 158 items are explicitly interaction-oriented (protocols, phrases, etiquette, practices). Early content emphasized descriptive facts about Oaxaca. Fieldwork redirected the cache toward knowledge that supports action: how to ask, order, greet, and participate appropriately.

2. Social context determines correctness

Literal accuracy proved insufficient. Participants emphasized that knowing what a term means is incomplete without knowing who can use it, with whom, and when.

Terms like paisano require explicit usage constraints. Knowing the definition is incomplete without knowing who can use it, with whom, and when.

LKC encodes this through structured metadata: adaptation_criticality (essential/important/helpful/enriching) and friction_without (offense/exploitation/misunderstanding/exclusion).

3. Visuals as knowledge

Images often communicated local meaning faster and more reliably than text. Visual reference material, particularly around food and everyday practices, emerged as a critical knowledge layer supporting recognition and embodied understanding.

1. Pilot participant review and rating workflows with local contributors
2. Refine verification and variant-handling mechanisms
3. Conduct usability testing focused on consultation scenarios
4. Explore offline and low-connectivity deployment contexts

LKC serves as the abstract knowledge layer connecting multiple projects:

• Learning from Oaxaca & AI-QA provided empirical constraints and analytic structure.
• Local Language Cards operationalized claims through embodied, paper-first interaction, feeding corrections back into the cache.

Andreev, H., Kosmas, P., Livieratos, A. D., Theocharous, A., & Zopiatis, A. (2025). Destination (Un)Known: Auditing bias and fairness in LLM-based travel recommendations. AI (Switzerland), 6(9), 236. https://doi.org/10.3390/ai6090236

Dudy, S., Tholeti, T., Ramachandranpillai, R., Ali, M., Li, T. J.-J., & Baeza-Yates, R. (2025). Unequal opportunities: Examining the bias in geographical recommendations by large language models. arXiv:2504.05325.

UNESCO. (2023). Guidance for generative AI in education and research. United Nations Educational, Scientific and Cultural Organization.