THE APPLICATION LAYER
Kompile App
A modular application framework with 40+ swappable modules, 44+ MCP tools, and 20+ data connectors. Knowledge graphs with MEBN probabilistic reasoning, causal inference, 9 maintenance primitives, and full MCP tool access for AI agents — swap any provider without touching business logic.
Modular by Design
Every capability is a separate module behind a shared interface. Swap any layer by changing a config — your pipeline code, agent logic, and API endpoints stay identical.
40+
Swappable Modules
44+
MCP Tools
20+
Data Connectors
9
Export Formats
The Crawl-to-Graph Pipeline
An 8-phase pipeline that ingests data, classifies content, extracts entities, builds knowledge graphs, resolves duplicates, and indexes vectors — with adaptive memory-aware parallelism at every step.
Pull data from 20+ sources (Confluence, Slack, web, S3, email, SQL) or load from local files. OAuth2 for cloud, robots.txt for web.
Auto-classify content: text PDFs, image-based PDFs, mixed, tables, audio, email. Route each to its specialized pipeline (text, visual, tables, email).
Ordered preprocessor chain: language detection, translation, boilerplate removal, Unicode normalization, PII redaction, content-hash + SimHash dedup.
Split into retrieval-sized passages with 5 strategies (sentence, recursive, markdown, token, table-aware). Register SNIPPET graph nodes with CONTAINS edges.
Multi-agent LLM + pattern-based entity extraction with cost-balanced batch planning. Schema enforcement (None/Lenient/Strict). Adaptive memory-aware parallelism.
Entity resolution via Levenshtein + embedding cosine + MEBN probabilistic scoring. Compute shared-entity and embedding-similarity edges across documents.
Generate vector embeddings (BGE, Arctic, SPLADE++) and store in Anserini HNSW, pgvector, Vespa, or Chroma. Hybrid BM25 + dense retrieval ready.
Post-crawl enrichment: deduplicate, prune, validate, normalize entities. Discover domain taxonomy via LLM, categorize entities, rebuild search indexes.
Knowledge Graphs, Reasoning & Agents
Kompile graphs are more than storage — they support probabilistic reasoning, causal inference, automated maintenance, and full MCP tool access for AI agents.
GraphRAG
Local ego-network, global community-level, and hybrid search modes with configurable vector weight. Three backends: JPA, Neo4j Cypher, or ND4J matrix. Louvain community detection with LLM-generated summaries.
Knowledge Graphs
Seven node levels, eleven edge types, named graph scoping, and multi-tenant fact sheet isolation. LLM or manual construction with schema enforcement (None/Lenient/Strict). Export to 9 formats including JSON-LD, GraphML, Cypher, and Obsidian.
Bayesian Networks
Multi-Entity Bayesian Networks (MEBN) with MFrag templates for EntityRelevance, CausalInfluence, InformationFlow, and RiskPropagation. Situation-specific grounding via BFS from seed nodes. Variable elimination for posterior inference.
Causal Inference
Eight W3C PROV-DM causal edge types: Causes, Enables, Triggers, Contributes To, Prevents, Correlates With, Influences, Derived From. Temporal chain extraction, attribution paths, and counterfactual modeling.
Graph Maintenance
Nine automated primitives: TTL sweep, orphan cleanup, confidence pruning, component pruning, contradiction detection, provenance validation, entity re-resolution, stats refresh, and community rebuild. Full mutation audit with before/after snapshots and WebSocket broadcasting.
MCP Graph Tools
30+ MCP tool operations: node/edge CRUD, bulk create, merge nodes, BFS traversal, ego networks, hybrid search, shortest path, PageRank, centrality, Louvain communities, and LLM community summaries — all callable by AI agents.
Graph Embeddings
Native TransE and RotatE knowledge graph embeddings trained with margin ranking loss and self-adversarial negative sampling. Link prediction, entity similarity, and head/tail prediction backed by ND4J tensors.
ReAct Agents & A2A
Reason-Act-Observe agents with pluggable interfaces. Agent-to-Agent (A2A) protocol at /.well-known/agent-card.json. KClaw browser-based agent runner with MCP injection and permission management.
Guardrails & Enforcer
Input guards (PII, injection, toxicity, topic filtering) and output guards (hallucination, relevancy, format). Enforcer wraps any agent with keyword, LLM, or subprocess judge modes for policy-governed execution.
Module Directory
Every layer of the stack is a swappable module. Pick what you need.
LLM Providers
Swap between hosted providers, local SameDiff models, or CLI-based agents. Every provider implements the same Spring AI interface — your code never changes.
OpenAIAnthropicGoogle GeminiSameDiff Local (SmolLM, Phi-2)CLI Agents (Claude, Codex, Gemini, Qwen)Embedding Models
Dense embeddings (BGE, Arctic, E5), sparse neural retrieval (SPLADE++, UniCoil), and cross-encoder rerankers (MiniLM, TinyBERT) — all running natively via SameDiff.
OpenAI EmbeddingsBGE / Arctic (SameDiff)SPLADE++ SparseSentence TransformersPostgresMLCross-Encoder RerankersVector Stores
Hybrid BM25 keyword + dense HNSW vector retrieval with Anserini, or connect to pgvector, Vespa, or ChromaDB. One API, four backends.
Anserini / Lucene HNSWPostgreSQL pgvectorVespaChromaDBDocument Loaders
PDF with table extraction, Office docs, Excel with formula dependency graphs, email inboxes (IMAP/POP3/MBOX/PST/EML), audio transcription, and BFS web crawling.
PDF + Tables (Tabula)Microsoft OfficeExcel + Formula GraphsEmail (IMAP, POP3, MBOX, PST)Audio (Whisper)Apache TikaWeb CrawlerData Sources
20+ connectors with OAuth2 authentication. Crawl SaaS tools, cloud storage, email, remote folders, and databases — all compiled into the graph automatically.
ConfluenceJiraNotionSlackDiscordGoogle WorkspaceOneDriveRedditGmailS3 / SFTP / SMBSQL DatabasesText Chunkers
Five pluggable strategies. Table-aware chunking keeps tables atomic and never splits them. Sentence detection supports 30+ languages via OpenNLP.
Sentence (OpenNLP)Recursive CharacterMarkdown-AwareToken-BasedTable-AwareKnowledge Graphs
Full knowledge graph stack with entity extraction, Bayesian reasoning (MEBN), causal inference, graph algorithms, TransE/RotatE embeddings, and export to JSON-LD, GraphML, Cypher, Obsidian, and more.
Neo4j (APOC)Native JPA + ND4J MatrixTransE / RotatE EmbeddingsMEBN / Bayesian NetworksCausal Inference9 Export FormatsMCP Tools & Agents
Expose 44+ tools via MCP stdio/SSE to Claude Code, Codex, Gemini CLI, and Qwen. Graph CRUD, traversal, algorithms, and community detection — all callable by AI agents natively.
44+ MCP ToolsGraph Mutation ToolsGraph Traversal ToolsGraph Algorithm ToolsAgent-to-Agent (A2A)ReAct Agent FrameworkKClaw Agent RunnerGuardrails & Evaluation
Input guards (PII, injection, toxicity, topic filtering) and output guards (hallucination, relevancy, format). Enforcer wraps any agent with keyword, LLM, or subprocess judge modes.
PII DetectionPrompt InjectionToxicity FilterHallucination DetectionRelevancy ScoringEnforcer (keyword/LLM/subprocess)Compute Graphs & Workflows
Orchestrate complex multi-step pipelines beyond simple chains. Plug in workflow engines, rule systems, or visual builders through the same Kompile interface.
Apache Camel
Enterprise integration patterns for routing, transformation, and mediation across 300+ connectors.
Drools
Business rules engine for decision logic that changes independently of application code.
n8n
Visual workflow automation for building complex multi-step data and AI pipelines.
Xircuits
Visual component-based workflows for ML pipelines with drag-and-drop composition.
Excel Workflows
Excel-driven workflows with formula dependency graphs extracted into executable compute graphs.
Scripting
Script-based workflows for custom pipeline logic with Python, Groovy, or JavaScript steps.
Ready to compile your stack?
Get early access and start building on the modular AI platform.
Request Early Access