~/wiki
Récent+ Ajouter

Incremental Knowledge Building

Confiance : high
incremental-knowledgeknowledge-integrationsynthesis-updatingcross-referencingllm-wikipersistent-learningwiki-maintenancearchitecture-patternsandrej-karpathycompounding-artifacts

The process of systematically adding new information to existing knowledge bases through integration rather than simple accumulation. Core methodology underlying compounding-artifacts and the llm-wiki-pattern. Contrasts sharply with traditional document storage approaches where each source exists in isolation.

Core Process

When new information arrives, incremental knowledge building involves:

  1. Reading and extraction: Parse source for key information, insights, entities
  2. Integration analysis: Identify how new information relates to existing knowledge
  3. Synthesis updating: Revise summaries and overviews to reflect new understanding
  4. Cross-referencing: Create links between new information and relevant existing pages
  5. Contradiction detection: Flag where new data challenges existing claims
  6. Entity maintenance: Update person, organization, concept pages with new details

Implementation in LLM Wikis

In andrej-karpathy's framework, incremental knowledge building happens during the ingest operation:

  • LLM reads new source document
  • Extracts key information and discusses takeaways
  • Writes summary page for the source
  • Updates index with new entry
  • Updates relevant entity and concept pages across wiki
  • Creates new pages for previously uncovered topics
  • Maintains cross-references and flags contradictions

Single source might touch 10-15 wiki pages during integration.

Contrast with Document Accumulation

Traditional approach (Document Libraries):

  • Add new document to collection
  • Document exists in isolation
  • No integration with existing knowledge
  • Connections must be discovered anew each time
  • Knowledge scattered across disconnected files

Incremental Knowledge Building:

  • New information integrates into existing structure
  • Cross-references automatically maintained
  • Contradictions flagged for resolution
  • Synthesis evolves to reflect growing understanding
  • Knowledge compounds rather than accumulates

Maintenance Challenge

The key insight: humans abandon knowledge bases because maintenance burden grows faster than value. Updating cross-references, keeping summaries current, noting when new data contradicts old claims - this bookkeeping is tedious but essential.

LLMs excel at incremental knowledge building because they:

  • Don't get bored with repetitive maintenance tasks
  • Can update multiple files in single operation
  • Maintain consistency across large collections
  • Don't forget to update cross-references

Quality Control

Effective incremental knowledge building requires:

  • Schema adherence: Consistent formatting and organization
  • Citation tracking: Clear source attribution for all claims
  • Contradiction flagging: Explicit noting of conflicting information
  • Quality assessment: Confidence ratings for different claims
  • Regular linting: Periodic health checks for consistency

Applications

Any domain where knowledge accumulates over time:

  • Research synthesis across multiple papers
  • Personal learning from diverse sources
  • Business intelligence aggregation
  • Competitive analysis updates
  • Course note integration

See also