54d9540c08
Files changed: - .wikitool-kb.json - AGENTS.md - CHANGES.md - INSTALL-MCP.md - INSTALL.md - README.md - VERSION - instructions/capture-session.md - instructions/dev/issue-tracking.md - instructions/german-terminology.md - instructions/kb-profiles.md - instructions/migrate-corpus.md - instructions/migrations/5.0.0-confidence-removal.md - instructions/private-instance.md - instructions/setup-instance.md - instructions/wiki-lint/SKILL.md - instructions/wiki-manage/SKILL.md - instructions/wiki-query/SKILL.md - kb/CONTRACT.md - kb/CONVENTIONS.md - kb/CONVENTIONS.md.template - kb/concepts/architectures/Consolidation Tiers.md - kb/concepts/architectures/Context Isolation.md - kb/concepts/architectures/Cross-platform Agent Skills.md - kb/concepts/architectures/Episodic Memory.md - kb/concepts/architectures/Hybrid Search.md - kb/concepts/architectures/Implementation Spectrum.md - kb/concepts/architectures/Knowledge Graph.md - kb/concepts/architectures/LLM Wiki Pattern.md - kb/concepts/architectures/MCP-Leseserver.md - kb/concepts/architectures/Memory Lifecycle.md - kb/concepts/architectures/OKF Compatibility.md - kb/concepts/architectures/Optional Instance Context File.md - kb/concepts/architectures/Personalization Plane.md - kb/concepts/architectures/Procedural Memory.md - kb/concepts/architectures/RAG.md - kb/concepts/architectures/Scale Ceiling.md - kb/concepts/architectures/Semantic Memory.md - kb/concepts/architectures/Three-Layer Architecture.md - kb/concepts/architectures/Token Economics.md - kb/concepts/architectures/Working Memory.md - kb/concepts/decisions/Delete Rather Than Anonymize.md - kb/concepts/decisions/Denylist over Allowlist.md - kb/concepts/decisions/Diff-Reviewable Agent Edits.md - kb/concepts/decisions/Dual Licensing by File Plan.md - kb/concepts/decisions/Issue Label Scheme.md - kb/concepts/decisions/KB Stack Versioning.md - kb/concepts/decisions/Structural Enforcement over Documented Rule.md - kb/concepts/patterns/Audit Trail.md - kb/concepts/patterns/BM25.md - kb/concepts/patterns/Command Round-Trip Integrity.md - kb/concepts/patterns/Confidence Scoring.md - kb/concepts/patterns/Contradiction Resolution.md - kb/concepts/patterns/Entity Extraction.md - kb/concepts/patterns/Filter on Ingest.md - kb/concepts/patterns/Forgetting.md - kb/concepts/patterns/Graph Traversal.md - kb/concepts/patterns/Mesh Sync.md - kb/concepts/patterns/Quality Scoring.md - kb/concepts/patterns/Reciprocal Rank Fusion.md - kb/concepts/patterns/Self-Healing.md - kb/concepts/patterns/Shared vs Private.md - kb/concepts/patterns/Typed Relationships.md - kb/concepts/patterns/Vector Search.md - kb/concepts/patterns/Work Coordination.md - kb/concepts/problems/Ambient Environment Dependency.md - kb/concepts/problems/Detect-Repair Asymmetry.md - kb/concepts/problems/Green Suite Blind Spot.md - kb/concepts/problems/Naming Convention Conflict.md - kb/concepts/problems/Write-Once Frontmatter Fields.md - kb/concepts/protocols/CPPC.md - kb/concepts/protocols/Modbus.md - kb/concepts/protocols/SSD TRIM.md - kb/concepts/workflows/Anti-Cramming Heuristic.md - kb/concepts/workflows/Bulk Operations.md - kb/concepts/workflows/CI Integration.md - kb/concepts/workflows/Checkpoint Audit.md - kb/concepts/workflows/Claude Code Auto Mode.md - kb/concepts/workflows/Content Quality Control.md - kb/concepts/workflows/Crystallization.md - kb/concepts/workflows/Event-Driven Automation.md - kb/concepts/workflows/Hooks.md - kb/concepts/workflows/Index Scaling.md - kb/concepts/workflows/Iteration and Cost Limits.md - kb/concepts/workflows/KB Migration.md - kb/concepts/workflows/Knowledge Compounding.md - kb/concepts/workflows/Lint Workflow.md - kb/concepts/workflows/Mass-Update Gate.md - kb/concepts/workflows/Multi-Agent Collaboration.md - kb/concepts/workflows/Privacy and Governance.md - kb/concepts/workflows/Publish-Remote Gate.md - kb/concepts/workflows/Quality and Self-Correction.md - kb/concepts/workflows/Semantic Lint Automation.md - kb/concepts/workflows/Session Orientation.md - kb/concepts/workflows/Split Merge Reclassify.md - kb/concepts/workflows/Split Threshold.md - kb/concepts/workflows/Stub Threshold.md - kb/concepts/workflows/Supersession.md - kb/concepts/workflows/User Management.md - kb/concepts/workflows/Workflow Extraction.md - kb/concepts/workflows/Workflow Orchestration.md - kb/entities/people/Andrej Karpathy.md - kb/entities/people/E3DC GmbH.md - kb/entities/people/Rohit Gupta.md - kb/entities/people/Vannevar Bush.md - kb/entities/projects/BCDModule.md - kb/entities/projects/Chemenu.md - kb/entities/projects/andybalholm-edl.md - kb/entities/projects/goresponsiveness.md - kb/entities/projects/ha-core.md - kb/entities/projects/hacs-e3dc.md - kb/entities/projects/hacs-integration-blueprint.md - kb/entities/projects/llm-wiki-skills.md - kb/entities/projects/plugnburn-edl.md - kb/entities/projects/wiki-skills-vanillaflava.md - kb/entities/projects/wiki-skills.md - kb/entities/systems/AGENTS.md.md - kb/entities/systems/CLAUDE.md.md - kb/entities/systems/E3DC.md - kb/entities/systems/ENVIRONMENT.md.md - kb/entities/systems/Memex.md - kb/entities/systems/Tolkien Gateway.md - kb/entities/technologies/Arch Linux.md - kb/entities/technologies/Disk Encryption.md - kb/entities/technologies/Docker.md - kb/entities/technologies/GRUB.md - kb/entities/technologies/Gitea Actions.md - kb/entities/technologies/Gitea.md - kb/entities/technologies/Go.md - kb/entities/technologies/Home Assistant.md - kb/entities/technologies/Kernel PM Governors.md - kb/entities/technologies/LVM.md - kb/entities/technologies/Linux Kernel.md - kb/entities/technologies/MQTT.md - kb/entities/technologies/OPC UA.md - kb/entities/technologies/Python.md - kb/entities/technologies/Rust.md - kb/entities/technologies/Wine GE.md - kb/entities/technologies/Wine-Staging.md - kb/entities/technologies/acpi-cpufreq.md - kb/entities/technologies/amd-pstate.md - kb/entities/technologies/iii Engine.md - kb/entities/tools/AUR.md - kb/entities/tools/Act Runner.md - kb/entities/tools/Agent Memory.md - kb/entities/tools/Aura.md - kb/entities/tools/Bottles.md - kb/entities/tools/ChatGPT.md - kb/entities/tools/Claude Code.md - kb/entities/tools/Codex CLI.md - kb/entities/tools/Dataview.md - kb/entities/tools/GPG.md - kb/entities/tools/GitHub Copilot.md - kb/entities/tools/Gitea MCP Server.md - kb/entities/tools/Lutris.md - kb/entities/tools/Marp.md - kb/entities/tools/Mistral Vibe.md - kb/entities/tools/NotebookLM.md - kb/entities/tools/Obsidian Web Clipper.md - kb/entities/tools/Obsidian.md - kb/entities/tools/OpenAI Codex.md - kb/entities/tools/OpenCode.md - kb/entities/tools/Pi.md - kb/entities/tools/Proton.md - kb/entities/tools/Steam.md - kb/entities/tools/Wine.md - kb/entities/tools/awesome-llm-wiki.md - kb/entities/tools/farzaa gist.md - kb/entities/tools/gdeploy.md - kb/entities/tools/makepkg.md - kb/entities/tools/pascalandy schema.md - kb/entities/tools/qmd.md - kb/entities/tools/wikitool.md - kb/index.md - kb/log.md - raw/CONTRACT.md - tools/CONTRACT.md - tools/README.md - tools/chemenu/api.py - tools/chemenu/cli.py - tools/chemenu/commands/confidence_decay.py - tools/chemenu/commands/docs_verify.py - tools/chemenu/commands/doctor.py - tools/chemenu/commands/index_build.py - tools/chemenu/commands/new_page.py - tools/chemenu/commands/search.py - tools/chemenu/commands/touch.py - tools/chemenu/commands/version_cmd.py - tools/chemenu/conventions.py - tools/chemenu/corpus_diff.py - tools/chemenu/frontmatter_io.py - tools/chemenu/lint_core.py - tools/chemenu/mcp/server.py - tools/chemenu/page.py - tools/chemenu/search/base.py - tools/chemenu/search/filters.py - tools/chemenu/search/ripgrep.py - tools/chemenu/search/service.py - tools/chemenu/search/types.py - tools/chemenu/tests/conftest.py - tools/chemenu/tests/test_api.py - tools/chemenu/tests/test_confidence_decay.py - tools/chemenu/tests/test_corpus_diff.py - tools/chemenu/tests/test_docs_verify.py - tools/chemenu/tests/test_frontmatter_io.py - tools/chemenu/tests/test_index_build.py - tools/chemenu/tests/test_kb_scan.py - tools/chemenu/tests/test_lint.py - tools/chemenu/tests/test_mcp_server.py - tools/chemenu/tests/test_new_page.py - tools/chemenu/tests/test_page_ops.py - tools/chemenu/tests/test_provenance.py - tools/chemenu/tests/test_raw_cmd.py - tools/chemenu/tests/test_search.py - tools/chemenu/tests/test_touch.py - tools/chemenu/tests/test_type_resolver.py - tools/chemenu/tests/test_version_cmd.py - tools/chemenu/tests/test_xref.py - tools/chemenu/version.py - types/concept.md - types/concept.schema.yaml - types/entity.md - types/entity.schema.yaml - types/instruction.md - types/type-spec.md
120 lines
4.3 KiB
Markdown
120 lines
4.3 KiB
Markdown
---
|
|
type: types/concept.md
|
|
concept_type: architecture
|
|
tags: [ai, retrieval, generation, knowledge-management]
|
|
created: 2026-07-26
|
|
modified: 2026-08-29
|
|
related:
|
|
- see-also: LLM Wiki Pattern
|
|
- see-also: NotebookLM
|
|
- see-also: ChatGPT
|
|
sources: [Source - LLM Wiki Pattern]
|
|
provenance: sourced
|
|
summary: Architekturmuster, bei dem LLMs die Generierung um Dokumente aus einer Wissensbasis anreichern.
|
|
---
|
|
# RAG
|
|
|
|
**Typ:** Architecture (Retrieval Augmented Generation)
|
|
|
|
## Definition
|
|
|
|
RAG (Retrieval Augmented Generation) ist ein KI-Architekturmuster, bei dem ein großes Sprachmodell (LLM) relevante Informationen aus einer Knowledge Base abruft, bevor es eine Antwort generiert. Dies ermöglicht dem LLM, Antworten bereitzustellen, die in externen Dokumenten verankert sind, anstatt sich ausschließlich auf seine Trainingsdaten zu verlassen.
|
|
|
|
## Kernpunkte
|
|
|
|
### Wie RAG funktioniert
|
|
|
|
1. **Indexierung**: Dokumente werden verarbeitet und für die Suche indexiert
|
|
2. **Abruf**: Bei jeder Abfrage werden relevante Chunks aus dem Index abgerufen
|
|
3. **Augmentation**: Abgerufene Chunks werden zum LLM-Kontext/Prompt hinzugefügt
|
|
4. **Generierung**: LLM generiert eine Antwort basierend auf seinem Wissen und den abgerufenen Chunks
|
|
|
|
### Traditionelle RAG-Systeme
|
|
|
|
Beispiele sind:
|
|
- [[NotebookLM]] (Google)
|
|
- [[ChatGPT]] Datei-Uploads (OpenAI)
|
|
- Die meisten kommerziellen RAG-Implementierungen
|
|
|
|
### Begrenzungen von traditionellem RAG
|
|
|
|
Laut dem [[LLM Wiki Pattern]]-Artikel:
|
|
|
|
1. **Keine Knowledge Accumulation**: Wissen wird bei jeder Abfrage von Grund auf neu abgeleitet
|
|
2. **Keine persistente Synthese**: Verbindungen zwischen Dokumenten werden nicht aufrechterhalten
|
|
3. **Keine Querverweise**: Keine expliziten Links zwischen verwandten Konzepten über Quellen hinweg
|
|
4. **Keine Widerspruchserkennung**: Konfliktinformationen werden nicht gekennzeichnet
|
|
5. **Kein Compounding**: Das Hinzufügen neuer Quellen baut nicht auf vorherigem Verständnis auf
|
|
6. **Ineffizient für komplexe Abfragen**: Subtile Fragen, die eine Synthese mehrerer Dokumente erfordern, müssen jedes Mal neu abgeleitet werden
|
|
|
|
### Wann RAG geeignet ist
|
|
|
|
- Bei einfachen, einmaligen Fragen
|
|
- Für schnelle Informationssuche
|
|
- Wenn Persistenz und Compounding nicht erforderlich sind
|
|
- Wenn der Aufwand für Wiki-Wartung nicht gerechtfertigt ist
|
|
|
|
### Wann über RAG hinausgehen
|
|
|
|
Das [[LLM Wiki Pattern]] in Betracht ziehen, wenn:
|
|
- Wissen im Laufe der Zeit angesammelt werden soll
|
|
- Persistente Querverweise benötigt werden
|
|
- Widersprüche zwischen Quellen gekennzeichnet werden sollen
|
|
- Wissen benötigt wird, das zusammengesetzt wird, wenn neue Quellen hinzugefügt werden
|
|
- Komplexe Abfragen, die Multi-Dokument-Synthese erfordern, häufig vorkommen
|
|
|
|
## RAG vs. LLM-Wiki-Pattern
|
|
|
|
| Aspekt | RAG | LLM-Wiki-Pattern |
|
|
|--------|-----|-------------------|
|
|
| Knowledge Accumulation | Nein | Ja |
|
|
| Persistente Querverweise | Nein | Ja |
|
|
| Widerspruchserkennung | Nein | Ja |
|
|
| Knowledge Compounding | Nein | Ja |
|
|
| Wartung | Automatisch | LLM-gepflegt |
|
|
| Abfrage-Geschwindigkeit | Schnell | Schnell (nach Kompilierung) |
|
|
| Setup-Komplexität | Niedrig | Mittel |
|
|
| Am besten für | Einmalige Abfragen | Laufende Knowledge Accumulation |
|
|
|
|
## Implementierungen
|
|
|
|
### RAG-Varianten
|
|
- Naive RAG: Einfache Ähnlichkeitssuche
|
|
- Vector RAG: Embedding-basierter Abruf
|
|
- Hybrid RAG: Kombiniert Schlüsselwort- und Vector-Suche
|
|
- Graph RAG: Verwendet Knowledge Graphs für Abruf
|
|
|
|
### LLM-Wiki-Pattern als verbessertes RAG
|
|
Das [[LLM Wiki Pattern]] kann als eine Verbesserung zu RAG angesehen werden, die hinzufügt:
|
|
- Persistenz-Schicht (das Wiki)
|
|
- Automatische Querverweise
|
|
- Widerspruchserkennung
|
|
- Knowledge Compounding
|
|
- Menschliche Kuratierung in der Schleife
|
|
|
|
## Tools, die RAG verwenden
|
|
|
|
- [[NotebookLM]]
|
|
- [[ChatGPT]] (mit Datei-Uploads)
|
|
- Viele unternehmensweite Knowledge-Management-Systeme
|
|
- Verschiedene Open-Source-RAG-Frameworks (LangChain, LlamaIndex, etc.)
|
|
|
|
## Geschichte
|
|
|
|
- [2020] - Frühe RAG-Papers und -Implementierungen
|
|
- [2023] - Kommerzielle RAG-Systeme entstehen (NotebookLM, ChatGPT Datei-Uploads)
|
|
- [2023-2024] - LLM-Wiki-Pattern entwickelt als Verbesserung zu RAG
|
|
- [2026-07-26] - Konzeptseite erstellt
|
|
|
|
## Siehe auch
|
|
|
|
- [[Knowledge Compounding]]
|
|
|
|
<!-- wikitool:links -->
|
|
## Beziehungen
|
|
|
|
- **see-also:** [[LLM Wiki Pattern]]
|
|
- **see-also:** [[NotebookLM]]
|
|
- **see-also:** [[ChatGPT]]
|
|
<!-- /wikitool:links -->
|