From Question toVerified Answer
For Decision Makers
Most AI tools match your question to similar text. That works for simple lookups — but falls apart on complex questions, multi-document reasoning, or anything requiring verification.
Courdx runs three retrieval strategies in parallel, validates every result, and cites every fact to the exact source sentence. You get answers you can trust — not answers you have to double-check.
For Technical Teams
The entire pipeline is orchestrated by 15+ LangGraph agentic workflows — not a monolithic prompt chain. Each stage is independently configurable, observable, and tuneable from the admin panel.
Semantic search (6 embedding models), Memgraph knowledge graph traversal, BM25 keyword matching, cross-encoder reranking, Corrective RAG, and RAGAS evaluation — all running in a production-grade pipeline.
The RAG Pipeline
Every Query. Five Stages, Nine Steps.
Nine steps grouped into five intelligent stages. Hover over each node to explore the pipeline — notice how three retrieval strategies branch in parallel, then converge through fusion and validation.
Data Flow
From Raw Document to Searchable Knowledge
Every document goes through ingestion, parsing, chunking, entity extraction, embedding, and indexing — fully automated.
Three Pillars of Intelligent Retrieval
Each pillar solves a fundamental problem that basic retrieval systems ignore.
Intelligent Retrieval
We Don't Just Search. We Understand.
Five layers of intelligence between your question and your answer: query decomposition breaks complex questions apart, HyDE predicts what ideal answers look like, multi-strategy search retrieves from vectors, graphs, and keywords simultaneously, cross-encoder reranking scores true relevance, and Corrective RAG self-corrects when confidence drops.
Knowledge Graph
See the Connections Others Miss
Courdx automatically extracts entities and relationships from your documents and stores them in a Memgraph knowledge graph. Community detection algorithms cluster related entities into groups — so you can ask big-picture questions like "what themes recur across our compliance reports?" that vector search can't touch.
Trust & Citations
Every Fact. Every Source. Verified.
Every claim cites the exact sentence in the source document — not just "found in Document 3." Confidence scores tell you how strong each citation is. Input and output guardrails block prompt injection, PII leaks, and hallucinated content. Full audit trail for compliance.
Multi-Dimensional Retrieval
Not Just Vector Search. Multi-Strategy Intelligence.
A single retrieval method always has blind spots. Vector search misses exact terms. Keyword search misses synonyms. Neither understands entity relationships.
Courdx runs all three simultaneously and fuses the results with Reciprocal Rank Fusion (RRF) — so you get the best of every approach in one ranked list.
Semantic Understanding
Vector embeddings capture meaning beyond keywords
Graph Reasoning
Memgraph knowledge graph reveals entity relationships
Keyword Precision
BM25 matching for codes, identifiers, and exact phrases
Query Intelligence
HyDE prediction + decomposition + smart routing
Step by Step
The Complete Journey
From natural language question to verified, cited answer — the five stages that power every query.
Your Question
Natural language query — plain English, no syntax required
Query Decomposition
Complex questions are broken into focused sub-queries. Each part gets its own retrieval pass.
- Predicts ideal answer shape using HyDE (Hypothetical Document Embeddings)
- Expands acronyms, synonyms, and domain-specific terms
- Routes each sub-query to the optimal retrieval strategy
Multi-Strategy Retrieval
Three search strategies run in parallel — each catching what the others miss.
- Vector search: semantic similarity via 6 embedding models including Qwen3, Nomic, MxBai, and Snowflake Arctic
- Knowledge graph: Memgraph traversal for entity relationships across documents
- BM25 keyword: precision matching for codes, names, identifiers, and exact phrases
- Reciprocal Rank Fusion (RRF) merges all results into one optimized ranking
Intelligent Validation
Every result is scored, reranked, and validated before it reaches the LLM.
- Cross-encoder model scores each query-document pair for true relevance (not just similarity)
- Low confidence triggers Corrective RAG — automatic query refinement and retry
- Falls back to web search if internal sources are insufficient
- RAGAS evaluation framework measures faithfulness, relevance, and answer quality
Cited, Verified Response
The final answer is synthesized from validated sources with mandatory citations.
- Every fact cited to the exact sentence in the source document
- Confidence score per citation — not just per answer
- Output guardrails check for hallucinations, PII leaks, and toxic content
- Full audit trail: who asked, what was retrieved, what was generated
Measurable Results
Up to 95% Accuracy. Measured.
Each retrieval strategy catches what the others miss. Combined with cross-encoder reranking and Corrective RAG, the multi-strategy approach closes the gaps that make basic AI search unreliable.
Verifiable Answers
Every Fact. Every Source. Verified.
Click any citation to see the exact source sentence highlighted in context. No more "the AI said so."
The company achieved $2.4M in Q3 revenue[1], representing a 23% YoY growth[2]. The main drivers were the new enterprise contracts signed in July and August[3].
Hover over highlighted text to see source
“$2.4M in Q3 revenue”
“23% YoY growth”
“enterprise contracts signed in July and August”
Every fact in Courdx responses is traced back to the exact sentence in your source documents
Observable Pipeline
Watch It Work. In Real Time.
Courdx isn't a black box. The admin panel includes 15 dedicated health pages showing real-time system metrics, retrieval quality, and cost analytics. You see exactly what the system is doing — and where to tune it.
Research-Backed
Built on Peer-Reviewed Research
Not prompt engineering. Real retrieval science, implemented in production.
GraphRAG Community Detection
Clusters related entities for big-picture questions
Corrective RAG
Self-healing retrieval that refines when confidence drops
Cross-Encoder Reranking
Deep relevance scoring, not just cosine similarity
Reciprocal Rank Fusion
Merges multi-strategy results into one optimal ranking
HyDE Predictive Embeddings
Generates ideal answer shape, then finds matching docs
Query Decomposition
Breaks complex questions into precisely answerable parts
LangGraph Orchestration
15+ agentic workflows coordinating retrieval and validation
RAGAS Evaluation
Automated quality measurement: faithfulness, relevance, accuracy
See the Pipeline In Action
Bring your hardest question and your real documents. We'll trace the entire retrieval pipeline live.