Architecture: ReshapeX Knowledge Grounding Layer [Skip to main content](#main-content)[Skip to navigation](#navigation)[Skip to footer](#footer) ARCHITECTURE # How the knowledge grounding layer actually works. Standard approaches ceiling near 80% on real industrial questions. The layer, the harness, and the evals are how you get the rest of the way to 99.9%. A graph-based structure of your catalog, specs, and docs, reached through a purpose-built tool harness, kept current with continuous sync. This is the architecture underneath every agent, and what it takes off your team’s plate. GRAPH · HARNESS · EVALS · SYNC * * * WHAT SITS UNDER YOUR AGENTS ## One verified layer underneath every agent. A knowledge grounding layer reviewed by your experts, reached through a purpose-built tool harness. Answering a questionBuilding the graph A question meets the Knowledge Graph ↺ Replay Incoming question Semantic Index · extracted constraints Straight shaftStainless steelHard chromeg6 toleranceØ 8 mmL = 200 mm Customer wording resolved to catalog attributes. No part number in the question. Knowledge Graph · traversal 0 candidate parts Grounded answer PSSFJ8-200 **Precision linear shaft, straight (no machining).** SUS440C stainless, hard chrome, Ø8 mm g6, L=200 mm. _✓_L=200 within series range 20–800 mm _✓_All 6 constraints satisfied _✓_Provenance: spec row + graph path _✓_Linked across the graph: compatible shaft support + collar The same question, without the graph ↺ Replay What happens when the exact same query goes to a frontier model instead. Frontier LLM · no grounding LLM aloneFor an 8 mm stainless shaft at 200 mm with hard chrome and g6, you are most likely looking for part SFU8-200 or similar. **Fluent, instant, and never checked.** The model has read the internet, not the catalog. It produces the most plausible-looking SKU, and plausible is where the risk lives. Frontier LLM · with web search Result · category landing pageResult · 400-page series PDFResult · site configurator (JavaScript, not renderable) LLM + searchI found the right series page, but the exact part number is generated by the site’s configurator, which I can’t execute. You’ll need to open it and select material, coating, tolerance and length yourself. **Honest, but not an answer.** Search finds the right neighborhood, then hands the real work back to your customer: open the configurator, pick material, coating, tolerance, hope you clicked right. LLM alone → **"SFU8-200 or similar" · a guess**LLM + search → **a link and instructions**ReshapeX on the graph → **PSSFJ8-200 · grounded** How the graph gets built ↺ Replay A brand’s raw knowledge 0sources · datasheets · websites · APIs · manuals · support tickets · CAD · ERP Agent 01 Source Discovery Explores everything the brand has: APIs, pages, files, references → KnowledgeMap · typed source registry, provenance on every entry Agent 02 Knowledge Cartographer Samples each source, extracts what actually exists SERIESPARTMATERIALDIMENSIONTOLERANCEFITS-WITHSUPERSEDES → EntityVocabulary Agent 03 Schema Modeling Infers the graph schema from the vocabulary ✓ Human approves the proposal Agent 04 Entity Harvesting Full paginated harvest of every approved source → Writes through the Graph Construction Harness **Lean orchestrator.** Passes run context; each agent reads upstream artifacts from the knowledge store, never from chat history. Run 1 · every run enriches the KnowledgeMap Brand knowledge graph 0nodes & relationships · provenance on every one Powered by ReshapeX · KCS: Knowledge Graph · Product Database · Semantic Index KNOWLEDGE CAPTURE ## Your experts' knowledge, captured once — answerable forever. The exceptions a twenty-year veteran keeps in their head are the first thing a model gets wrong and the first thing you lose when they retire. 1. ### Capture PDFs with merged cells, line cards with undocumented footnotes, price books with a decade of exceptions — structured into a verified knowledge graph, not guessed at by a general-purpose model. 2. ### Retain Drive configurations, fieldbus compatibility, firmware revisions, approved substitutes, discontinued parts: the hard-won answers stay in the layer when the person who knew them moves on. 3. ### Reuse Every agent (Reshape Sales, Reshape Engineer, Reshape Service, Reshape Maintenance) answers from the same captured layer, cites the catalog line or spec page it used, and escalates when confidence is low. Continuous sync keeps the captured knowledge current as your catalog, pricing, and specs change. THE RESULTS ## What the layer changes on the ground. For OEMs & industrial distributors 60–90%Labor cost savings 2–3×More qualified leads +15–20%Average order value For system integrators, manufacturers & plant operators 50–70%Faster first response 50%Faster RFQ answers THE PLATFORM ## Four products. One layer underneath. Reshape Service, Sales, Engineer and Maintenance run on one shared platform. Every answer they give comes from the same knowledge grounding layer, so none of them hallucinates a part number, a firmware revision or an approved substitute. Reshape does not ask your team to move. It reads from the ERP, catalog and document stores you already run, and answers in the inbox, CRM and Teams channels where the questions already arrive. * ### [Reshape Service](/en/product/agents/customer-support) AI customer service for distributors and OEMs. * ### [Reshape Engineer](/en/product/agents/application-engineer) AI applications engineer for distributors and OEMs. * ### [Reshape Sales](/en/product/agents/sales-concierge) AI inside sales for distributors and OEMs. * ### [Reshape Maintenance](/en/product/agents/field-operations) AI field maintenance for manufacturers. * ### Knowledge grounding The graph-based structure of your catalog, specs and docs that every product answers from, kept current with continuous sync. * ### GRID Resolves a full BOM or competitor list against your catalog: cross-references, substitutes, stock. * ### INSIGHT Every support request arrives with order history, datasheets and a drafted, cited reply attached. * ### Integrations Reads your ERP, catalog and document stores. Writes back to your CRM and service desk. * ### Channels Email, chat and Teams. The answer shows up where the question was asked. * ### Analytics What was asked, what was answered, and the source behind every answer. * ### MCP and API The layer and the products, reachable from your own systems. Standard approaches ceiling near 80% on real industrial questions. The layer, the harness and the evals are how every product gets to 99.9%. APPS ## Two applications built on the layer. [GRID](/en/product/grid) Applications Engineering Paste a thousand-line BOM or a competitor list. GRID resolves every line against your catalog: cross-references, substitutes, price, stock. * 01Cross-reference resolution * 02Substitute & compatibility checks * 03Price and stock lookups [INSIGHT](/en/product/desk) Customer Support Every request arrives with order history, pricing rules, datasheets, and a drafted answer already attached. * 01Order history & pricing rules * 02Attached datasheets * 03Drafted, citation-backed replies GRID GRID · BOM resolve Paste BOMResolvepaste-buffer.xlsx — Sheet1 #Pasted lineResolved against catalogStatus 011SFA619400R1051Pilot device · 22mm · cross-referencedIN STOCK 02MTR-PROT-9920XRMotor protection relay · 9–13 AIN STOCK 03PWR-DRV-4410Discontinued · approved substitute PWR-DRV-4415SUBSTITUTE 04FBUS-MOD-DP12Fieldbus module · PROFIBUS DP · firmware 2.4+IN STOCK 05CTL-XR-0071-BTwo catalog matches · protocol variant unconfirmedREVIEW 4 of 5 resolved · 1 flagged for review INSIGHT INSIGHT · Customer reply InboundFrom: buyer@midwest-oem.com Can you confirm a replacement for the drive on PO 44815? Draft replySend PWR-DRV-4415 is the approved substitute. It matches your firmware revision and ships from stock. Cited · datasheet rev 2.4, contract tier 2 Grounded context * Order history · PO 44815 * Pricing rules · contract tier 2 * Datasheet · PWR-DRV-4415 ![](/_next/image?url=%2Fimages%2Fhomepage%2Ftwo-paths-technicians-collaborating.jpg&w=3840&q=75&dpl=dpl_3Pkr97x2tJszZDs4V4XuyZBNto8J) Path A · MCP ## Bring your own agents. You already use Copilot, Claude, ChatGPT Enterprise, or your own AI team’s agents. We expose ReshapeX as an MCP server. Your agents call it. Accuracy jumps from a generic-AI ceiling to **99.9% on your catalog.** * CPMicrosoft Copilot * CLClaude * GPTChatGPT Enterprise * +Your custom agents * Your team owns the agent layer * ReshapeX exposed as a typed MCP server * Knowledge layer maintained by us, continuously * Live in weeks · no migration Path B · FDE ## We build the agents. You don’t have an AI team, or you want the full solution. Our forward-deployed engineers embed with your senior team and ship production agents: Reshape Sales, Reshape Engineer, Reshape Service, and Reshape Maintenance, running on the same grounded knowledge layer. * Reshape Sales * Reshape Engineer * Reshape Service * Reshape Maintenance * Forward-deployed engineer on site * Production agents in your stack, your domains * Built once, owned by you * Same 99.9% measured accuracy SAME FOUNDATION, EITHER WAY ### The ReshapeX product knowledge layer Your catalog · compatibility rules · approved substitutes · datasheets · proposal archive · pricing tiers 99.9% Measured accuracy TALK TO US ## Give us your twenty hardest questions. We’ll demo on your SKUs, run your evals, and cite every answer. * Real Examples * Working Demo * Your Data Be the team that defends every answer and ships AI in months, not years. [Schedule a meeting](/en/schedule) Talk to the agent first