RAG Systems & Semantic Search
Build a knowledge system that retrieves relevant evidence, respects access boundaries and explains its sources. We combine CogniGraph, Cognitive-OCR and CogniGraph Chunker with evaluation on your documents, covering ingestion, search, answer generation and operational handover.
Our RAG Toolchain
Three in-house Rust tools that cover the full RAG pipeline — from raw documents to graph-augmented retrieval.
CogniGraph
In-house · Rust Knowledge Graph Engine
Knowledge graph engine combining document storage, graph relationships, and vector embeddings in a single service. Pluggable backends (ArangoDB for production scale, embedded CozoDB for zero-config deployments). Four search modes: vector, semantic, hybrid BM25+RRF, and graph-augmented with multi-hop expansion. Sandboxed Lua scripting for server-side graph operations.
Cognitive-OCR
In-house · Cognition-Aware Document Extraction
Document extraction that analyzes page structure before choosing a processing strategy. Local layout and OCR models handle suitable pages, while vision-language models can process complex content. We evaluate extraction quality and model usage against representative documents rather than assuming the same savings for every file.
CogniGraph Chunker
In-house · Cognition-Aware Text Chunking
Four chunking strategies in one tool — fixed-size, delimiter-based, semantic, and cognition-aware with 8-signal boundary scoring (entity continuity, discourse structure, heading context, topic shift, orphan risk). Markdown-aware, 70+ languages, graph export for knowledge bases. Available as CLI, REST API, Python bindings, and Docker.
What is RAG?
Retrieval-Augmented Generation supplies relevant source material to an LLM. Retrieval and generation both need evaluation; missing or conflicting evidence should lead to a qualified answer or a request for more information.
Why It Matters
- Responses based on your actual documents and data
- Up-to-date information without model retraining
- Source references where supporting evidence is available
- Deployment and provider selection matched to your data boundary
Our Approach
- Graph-augmented retrieval, not just vector similarity
- Cognition-aware chunking that preserves document structure
- Hybrid search combining dense vectors with BM25 sparse retrieval
- Source attribution, retrieval evaluation and insufficient-evidence handling
Hybrid Search Architecture
We combine multiple retrieval strategies to maximize relevance — no single approach works for all content.
Dense Vectors (Semantic)
- Captures meaning and context beyond keywords
- Handles synonyms, paraphrasing, cross-language queries
- OpenAI, Ollama, ONNX, and Cloudflare embedding providers
Sparse Vectors (Lexical)
- Exact keyword and terminology matching
- Product codes, identifiers, domain-specific terms
- BM25 via ArangoSearch, SPLADE where applicable
Graph-Augmented Retrieval
- Semantic search finds seed documents
- Graph traversal expands through typed relationships
- Context-aware results that follow the shape of your knowledge
Vector Databases
We match the database to the problem — not the other way around.
ArangoDB
Our primary choice for complex projects
Graph + document + vector in one database. AQL for complex queries, Foxx Microservices for server-side logic, native clustering. When your data has relationships that matter for retrieval, ArangoDB is where we start.
Qdrant
When pure vector performance is critical
Rust-based vector similarity search with payload filtering, distributed deployment, and exceptional throughput. Our choice for high-volume retrieval workloads where graph relationships aren't the primary concern.
LanceDB
Embedded, serverless-friendly
Zero-config vector database that runs embedded in your application. Ideal for edge deployments, prototyping, and use cases where running a separate database server adds unnecessary complexity.
Others
Weaviate · AstraDB · CozoDB
Weaviate for AI-native workflows with built-in vectorization. AstraDB for global distribution on Cassandra. CozoDB as an embedded alternative via CogniGraph's pluggable backend. Chosen based on your scale and infrastructure.
Document Processing Pipeline
RAG quality depends on ingestion quality. Our pipeline handles the hard parts — format detection, layout understanding, and intelligent splitting.
Format Support
- PDF with layout detection, OCR, and table extraction
- DOCX and PPTX with formatting preservation
- Markdown roundtrip parsing
- Images via VLM description
- HTML, CSV, JSON, and structured data
Chunking Strategies
- Cognition-aware with 8-signal boundary scoring
- Semantic chunking via embedding similarity
- Fixed-size with delimiter-aware boundaries
- Markdown-aware (tables, code blocks stay atomic)
- Cross-chunk entity tracking and graph export
Advanced Retrieval
- Query expansion and HyDE (Hypothetical Document Embeddings)
- Cross-encoder reranking for precision
- Contextual compression of retrieved passages
- Self-query metadata filtering from natural language
- Reciprocal Rank Fusion across retrieval strategies
Proven Implementations
RAG systems we've built and deployed in production.
Pharmaceutical Research
Fortune 500 — Clinical Document Intelligence
Knowledge-graph and RAG workflows for qualitative pharmaceutical research, document exploration and evidence-linked insights. Data access, retention and review requirements are agreed for each engagement; regulatory obligations depend on the complete system and its use.
WolfGPT
WOLF GmbH — Enterprise Knowledge System
Multi-LLM knowledge system with intent analysis and dynamic routing. RAG over company documentation with tool-calling for data retrieval and complex document processing workflows.
TranscriptIntel
In-house — Conversation Intelligence
RAG-powered analysis of interview transcripts — speaker identification, behavioral pattern extraction, and executive report generation with evidence trails across thousands of documents.
PLU Finder
Incubated — Produce Intelligence
Knowledge base of 1,540+ PLU codes with AI-powered semantic search and MCP server integration. Retrieval-augmented answers about storage guidelines, nutritional data, and produce identification.
Evaluation & Research
Measure retrieval and answer quality before expanding access or automating decisions.
Retrieval Acceptance
Project Deliverables — Evidence & Operations
We define representative questions, assess retrieval relevance and answer support, test permission boundaries and document update/deletion behavior. Delivery includes an ingestion plan, source references, evaluation results and operational guidance for freshness and fallback behavior.
CogniGraph Myelin
In-house Research — Local Relation Verification
Myelin checks candidate entity pairs against source text for relation, direction and assertion status. This research supports evidence-aware graph construction. It does not replace extraction, graph authorization or human review, and is not advertised as a released production integration.
Ready to build your AI-powered knowledge base?
Contact us to discuss how RAG — built on tools we wrote ourselves — can transform how your organization accesses and utilizes information.
Get in Touch