Bot:Guide/search

This page provides the bot with guidance on how to interpret page titles from user requests and how to search the wiki when an exact match is not found. The bot reads this page automatically at startup and on every polling cycle.

~*~

Naming convention

Pages on this wiki follow a structured naming convention using namespaces, subpages, and natural language titles.

Namespaces are prefixed before a colon. Common namespaces include:

• Web: — web-sourced content, organised by company and topic (e.g. Web:AXA/results/FY25/Press release summary)
• Definition: — glossary and terminology entries (e.g. Definition:Combined ratio)
• Bot: — bot configuration and service logs (internal use only)
• Prompt: — prompt templates (internal use only)
• Main namespace — pages without a prefix (e.g. Thomas Buberl, Cyber insurtech MGAs and underwriting agencies)

Subpages are separated by forward slashes /. They represent a hierarchy within a namespace or page. For example:

• Web:AXA/results/FY25/Press release summary — the page "Press release summary" is a subpage under Web:AXA/results/FY25

Interpreting user requests:

• When a user says "page X in Namespace:Path", the full title is Namespace:Path/X
• When a user says "page X in Path", look for Path/X or identify the most likely namespace
• When a user gives a plain name like "Thomas Buberl", the page is in the main namespace
• Page titles are case-sensitive for the first character after the namespace; subsequent words follow the title as stored on the wiki

Examples:

• "press release summary in Web:AXA/results/FY25" → Web:AXA/results/FY25/Press release summary
• "combined ratio in Definition" → Definition:Combined ratio
• "Thomas Buberl" → Thomas Buberl
• "the page FY25 results summary under Web:AXA" → Web:AXA/FY25 results summary

~*~

Intent routing

When the user's request implies a specific category of information, the bot should prioritise searching within the corresponding path:

User intent	Path	Examples
Contact information, team members, org structure	Internal:Teams/	"who is in Group tax", "contact for corporate finance"
Training materials, courses, learning	Internal:Training/	"the training on cyber risk", "onboarding course"
AXA financial results	Web:AXA/results/	"FY25 press release", "AXA half-year results"
Market intelligence, industry trends, sector watch	Web:Watch/	"M&A trend in insurance", "AI in insurance"
Country-specific insurance market data, market share, rankings	Web:Watch/{Country}/	"France market share", "market ranking in Germany", "insurance market in Japan"
Glossary, definitions, terminology	Definition:	"what is combined ratio", "define loss ratio"

This table is a hint, not a hard rule. If no result is found in the suggested path, the bot should fall back to a full wiki search.

Country market monitoring: Pages under Web:Watch/{Country}/ track the insurance market in that country. For example, Web:Watch/France/Market ranking contains market share and ranking data for France. When the user asks about a specific country's market, replace {Country} with the country name in the suggested path (e.g. Web:Watch/France/, Web:Watch/Germany/).

When the user asks for an overview, outline, or summary of an entire topic (e.g. "overview of IFRS17 training", "what topics does the cyber course cover"), include the root page of the path in page_hints. For example, if the path is Internal:Training/IFRS17/, add Internal:Training/IFRS17 to page_hints. This ensures the index/overview page is checked alongside subpages.

~*~

Search

When the bot cannot find an exact match for a requested page title, it performs a search across the wiki.

Search scope:

• The bot searches across all namespaces on the wiki
• The following namespaces are excluded from search results:
  • Bot: (namespace ID 3000) — internal bot configuration and logs
  • Prompt: (namespace ID 4002) — internal prompt templates
• These namespaces are excluded because they contain internal operational content that users would never want as a PDF

Search results:

• The bot retrieves the top 10 candidate pages from the search results
• An AI model then evaluates each candidate against the user's original request and selects the best match
• If no candidate is relevant, the bot reports that the page could not be found
• If a match is found but is not exact, the bot explains clearly in its reply that the attached PDF is the closest match, not the exact page requested

Confidence levels:

• High — the candidate clearly corresponds to the user's request (e.g. minor spelling difference or capitalisation mismatch)
• Medium — the candidate is likely correct but there is some ambiguity
• Low — the candidate is a stretch; the bot attaches it but warns the user to verify