agent

const (
	// DefaultAgentTemperature is the default temperature for the agent
	DefaultAgentTemperature = 0.7
	// DefaultAgentMaxIterations is the default maximum number of iterations for the agent
	DefaultAgentMaxIterations = 20
	// DefaultUseCustomSystemPrompt is the default whether to use custom system prompt for the agent
	DefaultUseCustomSystemPrompt = false
)

const (
	WikiGranularityGuidanceFocused = `` /* 1244-byte string literal not displayed */

	WikiGranularityGuidanceStandard = `` /* 1045-byte string literal not displayed */

	WikiGranularityGuidanceExhaustive = `` /* 800-byte string literal not displayed */

)

Granularity guidance blocks injected into WikiCandidateSlugPrompt. The pipeline resolves a KnowledgeBase's configured granularity to one of these strings via WikiGranularityGuidance().

The three levels form a spectrum from "only the document's main subjects" to "every named thing you see". Moving down the list monotonically increases the candidate slug count, the downstream chunk-citation cost, and the noise-to-signal ratio of the wiki index.

const WikiCandidateSlugPrompt = `` /* 4877-byte string literal not displayed */

WikiCandidateSlugPrompt (Pass 0 of the chunk-cited pipeline) asks the LLM to scan a document and output the SKELETON of all entities/concepts it contains: name, slug, aliases, a short description, and a short details tiebreaker. The heavy lifting — linking each slug to concrete supporting chunks — is done in a second pass (see WikiChunkCitationPrompt). Because this prompt no longer has to carry full facts per item, it stays cheap even for long docs.

const WikiChunkCitationPrompt = `` /* 2431-byte string literal not displayed */

WikiChunkCitationPrompt (Pass 1..N of the chunk-cited pipeline) asks the LLM to read a batch of chunks and, for each candidate entity/concept, list the chunk IDs that substantively discuss it. This keeps per-slug "facts" in their verbatim form (the chunk text) instead of asking the LLM to paraphrase. Block order matters for provider prefix caching: the static rules, output schema and the per-document-stable <candidate_slugs> are placed BEFORE the per-batch <chunks> block. Within one document only ChunksXML changes between batches, so every batch after the first shares the long [rules | candidate_slugs] prefix and avoids re-billing the static rules.

const WikiDeduplicationPrompt = `` /* 5555-byte string literal not displayed */

WikiDeduplicationPrompt asks the LLM to identify duplicate entities/concepts between newly extracted items and existing wiki pages.

const WikiIndexIntroPrompt = `` /* 607-byte string literal not displayed */

WikiIndexIntroPrompt generates the introduction for a NEW index page (first time only).

const WikiIndexIntroUpdatePrompt = `` /* 850-byte string literal not displayed */

WikiIndexIntroUpdatePrompt incrementally updates an existing index introduction.

const WikiKnowledgeExtractPrompt = `` /* 5532-byte string literal not displayed */

WikiKnowledgeExtractPrompt extracts both entities and concepts in a single LLM call. Returns a JSON object with "entities" and "concepts" arrays. This replaces the former separate WikiEntityExtractPrompt and WikiConceptExtractPrompt.

const WikiPageModifyContinuationPrompt = `Your previous message was cut off because it reached the ` +
	`output length limit, so the page is incomplete.

Continue the SAME message from exactly where it stopped:
1. Do NOT repeat anything you already wrote — not the SUMMARY line, not the headings, not the ` +
	`rows, not the opening of the row you were in the middle of.
2. Your reply is appended verbatim to what you already wrote. Its first characters must be exactly ` +
	`the characters that follow the last character of your previous message. If you stopped in the ` +
	`middle of a Markdown table row, finish that row first.
3. Keep the same structure, formatting, ordering, and language as the part you already wrote.
4. Output the continuation ONLY: no preamble, no apology, no explanation, no closing code fence.
5. If the page was in fact already complete, reply with exactly: (complete)`

WikiPageModifyContinuationPrompt resumes a page rewrite that the provider cut off at the completion budget (finish_reason=length). Long enumerations are the shape that hits the cap: a certificate ledger with a hundred-plus holder rows is emitted row by row, and the provider stops the model mid-table, so the page is persisted one third short. The partial page is replayed as an assistant turn and this prompt asks for the remainder — WITHOUT the replay the model answers the original question again from the top, which is the #3446 spiral the agent loop works around.

const WikiPageModifySystemPrompt = `` /* 3148-byte string literal not displayed */

WikiPageModifySystemPrompt contains only rules shared by every page update. Keeping page identity and source data out of this message gives providers a long byte-stable prefix to cache across a reduce batch.

const WikiPageModifyUserPrompt = `` /* 3143-byte string literal not displayed */

WikiPageModifyUserPrompt contains the per-batch and per-page data. The document- level source context deliberately comes first: all pages generated from one source then share the longest possible prefix before page metadata diverges.

const WikiSummaryPrompt = `` /* 2086-byte string literal not displayed */

WikiSummaryPrompt generates a summary page for a newly ingested document.

Filename and title are intentionally NOT passed to the LLM: documents uploaded to WeKnora often carry filenames that say nothing about the content (e.g. scanned PDFs named after the scanner model "MX5280.pdf"), and feeding such filenames to the model invites hallucinated summaries when the actual extracted content is thin. The model must rely solely on the document content provided below.

const WikiTaxonomyPlanPrompt = `` /* 2303-byte string literal not displayed */

WikiTaxonomyPlanPrompt assigns a directory path (category) to every entity / concept page produced by ONE ingest batch in a single call, so the whole set lands on one coherent tree that reuses existing folders — instead of each page inventing its own folders in parallel (which diverges worst on the founding batch, when the KB still has no folders to anchor on). The result is applied in reduce only to pages that don't already have a category, so user edits and previously-filed pages are never churned.