Skip to content

LLM Wiki is a cross-platform desktop application that turns your documents into an organized, interlinked knowledge base — automatically. Instead of traditional RAG (retrieve-and-answer from scratch every time), the LLM incrementally builds and maintains a persistent wiki from your sources。

TypeScript

fix: ground generated entity pages

nash_su773587db568d
diff --git a/src/lib/ingest.prompt.test.ts b/src/lib/ingest.prompt.test.ts
index 7be7f884..8b72bc97 100644
--- a/src/lib/ingest.prompt.test.ts
+++ b/src/lib/ingest.prompt.test.ts
@@ -2,6 +2,7 @@ import { describe, it, expect, beforeEach } from "vitest"
import {
buildReviewSuggestionPrompt,
buildAnalysisPrompt,
+ buildChunkAnalysisSystemPrompt,
buildGenerationPrompt,
buildPageMergeSystemPrompt,
computeIngestGenerationMaxTokens,
@@ -57,6 +58,19 @@ describe("buildAnalysisPrompt language directive", () => {
expect(prompt).toContain("Do not transfer claims, limits, or evaluations")
})
+ it("applies a significance and canonical-name threshold to entities", () => {
+ const prompt = buildAnalysisPrompt("", "", "")
+
+ expect(prompt).toContain("standalone-page significance threshold")
+ expect(prompt).toContain("passing mentions")
+ expect(prompt).toContain("ordinary story props")
+ expect(prompt).toContain("unnamed or generic-role people")
+ expect(prompt).toContain("Page-worthy: yes")
+ expect(prompt).toContain("reuse the exact title of an existing wiki page first")
+ expect(prompt).toContain("never invent or phonetically guess a name")
+ expect(prompt).not.toContain("central vs. peripheral")
+ })
+
it("injects the project schema so analysis can recommend custom-typed pages", () => {
const schema = "## Page Types\n| goal | wiki/goals/ | Outcomes |\n| habit | wiki/habits/ | Behaviours |"
const prompt = buildAnalysisPrompt("", "", "", schema)
@@ -144,6 +158,34 @@ describe("buildGenerationPrompt language directive", () => {
expect(prompt).toContain("cite which source/frontmatter `sources` entry supports that statement")
})
+ it("requires standalone entity pages to establish identity without promoting incidental mentions", () => {
+ const prompt = buildGenerationPrompt("", "", "", "source.pdf")
+
+ expect(prompt).toContain("only for a candidate explicitly marked `Page-worthy: yes`")
+ expect(prompt).toContain("wikilinks only when the target already exists")
+ expect(prompt).toContain("opening paragraph of every entity page must answer what the subject is")
+ expect(prompt).toContain("stable, high-confidence public knowledge")
+ expect(prompt).toContain("independently verifiable background in the mandatory output language")
+ expect(prompt).toContain("do not create the standalone page")
+ expect(prompt).toContain("reuse the exact title of an existing wiki page first")
+ expect(prompt).toContain("Final Global Digest is authoritative")
+ })
+
+ it("defers long-source entity significance decisions to the global digest", () => {
+ const prompt = buildChunkAnalysisSystemPrompt("", "", "", "long source")
+ const chunkStart = prompt.indexOf("## Chunk Analysis")
+ const digestStart = prompt.indexOf("## Updated Global Digest")
+ const entityRules = prompt.indexOf("Entity handling rules:")
+
+ expect(chunkStart).toBeGreaterThanOrEqual(0)
+ expect(digestStart).toBeGreaterThan(chunkStart)
+ expect(entityRules).toBeGreaterThan(digestStart)
+ expect(prompt).toContain("without making a final standalone-page decision from one chunk alone")
+ expect(prompt).toContain("accumulate recurrence across chunks")
+ expect(prompt).toContain("supporting chunks")
+ expect(prompt).toContain("Page-worthy: yes")
+ })
+
it("makes project schema routing authoritative over default entity and concept folders", () => {
const prompt = buildGenerationPrompt(
"Use wiki/people/ for people. Use wiki/technologies/ for technical methods.",
diff --git a/src/lib/ingest.ts b/src/lib/ingest.ts
index ddd8f3f7..5e2d33aa 100644
--- a/src/lib/ingest.ts
+++ b/src/lib/ingest.ts
@@ -2208,6 +2208,26 @@ function shouldRunDedicatedReviewStage(generation: string): boolean {
|| /---REVIEW:\s*[\w-]+\s*\|[\s\S]*$/i.test(generation)
}
+const ENTITY_ANALYSIS_RULES = [
+ "Apply a standalone-page significance threshold: page-worthy subjects are central to the source, recur meaningfully, carry substantive claims, or are reusable across the wiki.",
+ "Do not mark passing mentions, ordinary story props, unnamed or generic-role people, chapter/part labels, or internal workflow artifacts as page-worthy unless the source makes them independently important knowledge subjects.",
+ "For every candidate, output `Page-worthy: yes` or `Page-worthy: no` plus one short justification. Keep non-page-worthy candidates only as context for the source summary.",
+ "Canonical naming priority: reuse the exact title of an existing wiki page first; otherwise preserve the source spelling. Widely used localized names may be recorded as aliases, but never invent or phonetically guess a name.",
+] as const
+
+const ENTITY_PAGE_RULES = [
+ "Create a standalone entity page only for a candidate explicitly marked `Page-worthy: yes`. Keep other mentions as plain prose in the source summary, or as wikilinks only when the target already exists in the current wiki index.",
+ "Canonical naming priority: reuse the exact title of an existing wiki page first; otherwise preserve the source spelling. A widely used localized name may be an alias, but never invent or phonetically guess a name.",
+ "The opening paragraph of every entity page must answer what the subject is, not merely describe its role in this source.",
+ "Prefer an identity supported by the source or existing wiki. If one short identity sentence must use stable, high-confidence public knowledge to identify a well-known subject, clearly label it as independently verifiable background in the mandatory output language so it is not attributed to the imported source. Never add uncertain details; when neither the source nor existing wiki establishes an unambiguous identity, do not create the standalone page.",
+] as const
+
+const LONG_SOURCE_ENTITY_RULES = [
+ "In each Chunk Analysis, record named-subject mentions and identity evidence without making a final standalone-page decision from one chunk alone.",
+ "In Updated Global Digest > Entities, accumulate recurrence across chunks and output for each candidate: canonical name, identity, supporting chunks, significance evidence, and `Page-worthy: yes` or `Page-worthy: no`.",
+ ...ENTITY_ANALYSIS_RULES,
+] as const
+
/**
* Step 1 prompt: AI reads the source and produces a structured analysis.
* This is the "discussion" step — the AI reasons about the source before writing wiki pages.
@@ -2227,10 +2247,11 @@ export function buildAnalysisPrompt(
"Your analysis should cover:",
"",
"## Key Entities",
- "List people, organizations, products, datasets, tools mentioned. For each:",
+ "List candidate named people, organizations, products, datasets, and tools. For each:",
"- Name and type",
- "- Role in the source (central vs. peripheral)",
+ "- Role and significance in the source",
"- Whether it likely already exists in the wiki (check the index)",
+ ...ENTITY_ANALYSIS_RULES,
"",
"## Key Concepts",
"List theories, methods, techniques, phenomena. For each:",
@@ -2317,6 +2338,8 @@ export function buildGenerationPrompt(
"2. Entity or schema-defined typed pages for key named things identified in the analysis. Prefer schema-defined directories when present; otherwise use wiki/entities/.",
"3. Concept or schema-defined typed pages for key ideas, methods, techniques, and abstractions. Prefer schema-defined directories when present; otherwise use wiki/concepts/.",
"4. A log entry for wiki/log.md (just the new entry to append, format: ## [YYYY-MM-DD] ingest | Title)",
+ ...ENTITY_PAGE_RULES,
+ "For consolidated long-document analyses, the Final Global Digest is authoritative for `Page-worthy` decisions; per-chunk mention notes are evidence, not page-creation instructions.",
"Do not generate wiki/index.md or wiki/overview.md. The application maintains aggregate navigation separately so large wikis are never rewritten through model output.",
"",
"## Frontmatter Rules (CRITICAL — parser is strict)",
@@ -2873,7 +2896,7 @@ function extractMarkedSection(raw: string, heading: string): string {
return re.exec(raw)?.[1]?.trim() ?? ""
}
-function buildChunkAnalysisSystemPrompt(
+export function buildChunkAnalysisSystemPrompt(
purpose: string,
schema: string,
index: string,
@@ -2901,6 +2924,8 @@ function buildChunkAnalysisSystemPrompt(
"## Updated Global Digest",
"A compact document-level digest that incorporates this chunk and preserves prior cross-chunk context.",
"Keep this digest structured under: Summary, Entities, Concepts, Schema-Typed Candidates, Claims, Evidence, Contradictions, Open Questions, Cross-Chunk Relations.",
+ "Entity handling rules:",
+ ...LONG_SOURCE_ENTITY_RULES.map((rule) => `- ${rule}`),
"Use schema-defined types only when the source actually supports them; never invent goals, habits, journal entries, decisions, or similar user-authored records that are not present in the source.",
"",
"Stable project context follows. It changes rarely and should be treated as background:",