Axion is not a chatbot bolted onto a file server. It is a purpose-built intelligence layer that reads, classifies, connects, and cites every answer from your actual project documents. Here is how the system works, from upload to cited answer.
Upload a 1,200-page bid package. Axion parses every page: OCR for scanned documents, text extraction for digital files, metadata identification for revision tracking. The system processes specifications, contracts, RFIs, submittals, geotechnical reports, environmental studies, change orders, and 150 other document types native to solar, BESS, and construction projects.
A typical bid package is fully ingested in under ten minutes. Supported formats include PDF, DOCX, XLSX, TXT, CSV, and other text-based files. Native drawing ingestion is on the roadmap.
Axion builds a knowledge graph using Neo4j, a graph database designed for connected data. Instead of storing your documents as isolated files in folders, Axion extracts entities (equipment, specifications, parties, dates, requirements, locations) and maps the relationships between them.
A change order references a spec section that references an equipment schedule that references a submittal. In a folder, those connections are invisible. In the knowledge graph, they are explicit and queryable.
The graph is structured around a domain-specific ontology built from nearly 20 years of utility-scale solar and BESS project operations: 27 document categories, 52 work packages, 156 document types. This is not a generic construction taxonomy adapted from commercial building. It was built for your projects.
Every passage in your documents is embedded into a Pinecone vector store, a high-performance index that maps the meaning of text, not just the keywords. When you ask "Does the Division 26 spec require arc-flash labels on every disconnect?" the vector store finds the passages that answer that question, even if the spec uses different terminology than your query.
The knowledge graph and vector store work together. The graph narrows context to the right project, the right document set, the right revision. The vector store finds the most relevant passages within that context. Neither alone is as accurate as both together.
When you ask Axion a question, three things happen in sequence. First, the knowledge graph identifies which documents, entities, and relationships are relevant to your query. Second, the vector store retrieves the most semantically similar passages from that narrowed context. Third, a large language model generates a natural-language answer grounded entirely in the retrieved passages.
Every claim in the answer is cited to a specific document, section, and page. If the answer is not in your documents, Axion says so. It does not fabricate contract terms, invent risk flags, or hallucinate equipment specifications.
Every Axion deployment runs on dedicated, single-tenant infrastructure. Your knowledge graph is a separate Neo4j instance. Your vector store is a separate Pinecone namespace. Your encryption keys are separate. Your project data never enters another customer's reasoning context, and it is never used to train, fine-tune, or improve any AI model.
This is not a policy toggle in a multi-tenant system. It is a separate infrastructure deployment for every customer. We call it IP Fortress because your intellectual property, your contract positions, your cost data, and your risk assessments deserve the same isolation as your financials.
Each customer deployment uses its own cryptographic keys. Your data at rest and in transit is isolated at the encryption layer.
Your documents do not train any AI model. Not ours, not our providers'. This is a contractual commitment in the Terms of Service.
Your Neo4j knowledge graph and Pinecone vector store are separate instances. No shared databases, no shared reasoning context.
Export your outputs and request deletion at any time, with certification. You built the knowledge base. You own it.
The best way to evaluate Axion is with your own documents. Upload a real bid package, ask a real question, and see the cited answer. No demo data. No slides. Your documents, your questions, your evaluation.