Skip to main content

Calcifer

Overview​

Calcifer is a modular, containerized agentic AI platform and orchestrator designed for flexibility and ease of integration. Built on a microservices architecture with plug‑and‑play components, it provides advanced Retrieval‑Augmented Generation (GraphRAG), agentic AI workflows, and extensible tool integration via the Model Context Protocol (MCP).

We built Calcifer to meet many enterprise situations — cloud, hybrid, or fully self‑hosted — making advanced AI capabilities accessible and adaptable from day one.

Calcifer is Digital Tvilling’s high‑trust AI Orchestrator, designed to extract information and insights from all kinds of structured and unstructured data while maintaining traceability to source documents or data. Combining tried‑and‑tested approaches from computational linguistics and information science with recent developments in large language models, Calcifer can be used to set up information‑extraction pipelines, knowledge graphs, and agentic systems for natural‑language interaction and exploration.

With its modular and flexible design, Calcifer is easily adapted to fit specific contexts — from intelligence analysis and regulated domains to complex business operations. It works with a variety of local and cloud‑based models; for sensitive applications, it can run fully on a local server (or even a single computer) with detailed logging.

Key capabilities​

  • GraphRAG: unify graph analytics/traversals and text retrieval for context‑rich reasoning
  • Pipelining: ingest and process diverse sources into storage backends
  • Entity extraction and enrichment: NLP pipelines with results written straight into graph structures
  • Agentic orchestration: build workflows where agents call graphs, data lakes, and APIs
  • Extensible tools: integrate via MCP and standardized APIs

Deployment options​

  • Cloud, on‑prem, or hybrid deployments
  • Modular services with Docker/Kubernetes for isolation and scaling
  • Optionally, source code access under contract for sovereign deployments

Integrations​

  • Data backends: Neo4j (graph), S3 (object), PostgreSQL (relational)
  • Orchestration frameworks: LangChain/LangGraph
  • APIs & services: FastAPI for service exposure; MCP servers for external tool connectivity

Built‑in tools​

  • Pipeline management: configure, run, and monitor AI pipelines with control over each step
  • Model‑agnostic: swap or update AI models without rewriting entire workflows
  • Drag‑and‑drop files: easy file ingestion
  • Named Entity Recognition (NER): extract dates, names, organizations, and custom entities
  • Keyword extraction: lists, statistical techniques, or LLM‑based
  • Timeline generation: automatically build timelines from document sources
  • Integration with existing databases: use Calcifer alongside existing stores — no migration required
  • Cloud or on‑prem: fully operable in isolated or secure environments
  • Source traceability: see exactly which file/page/timestamp supported each output

With LLMs we also enable​

  • Summaries and queries: generate readable summaries or ask natural‑language questions
  • Embeddings: turn texts into vectors for clustering, similarity, and retrieval
  • RAG (Retrieval‑Augmented Generation): combine internal knowledge sources with LLM responses
  • Chat interfaces: embed Q&A or assistant‑style interactions in apps
  • Custom causal models for explanations: plug in domain‑specific models for interpretable, context‑aware explanations

Design principles​

  • Modularity: microservices with independent, containerized components
  • Traceability: AI outputs link back to exact source pages for verification
  • Flexibility: configurable, extensible, adaptable to diverse use cases
  • Plug‑and‑Play: standardized interfaces and protocols
  • Ease of integration: pre‑built components and standard APIs

Technical stack (at a glance)​

  • Languages: Python, JavaScript
  • Frameworks: LangGraph/LangChain; spaCy and transformer models for NLP
  • Data backends: Neo4j, S3, PostgreSQL
  • Packaging: Docker; deployable to K8s