Skip to content

Libre/ERP/PLM

Status: Active test environment; ERP/PLM search integration deployed
Deployment last verified: 2026-07-29
Documentation updated: 2026-07-29
Application: chat.rapiddraft.ai
Railway: bk-RAG-test / testing

This surface documents the connected engineering-information environment: LibreChat is the conversational workspace, Odoo is the operational ERP authority, Aras is the engineering and lifecycle PLM authority, and the RapidDraft integration services provide governed search and analysis paths. The systems remain independently authoritative; LibreChat presents results but does not become a system of record.

LibreChat and its databases run in Railway. Chat and embedding inference run on the RapidDraft Local AI Server and reach Railway through authenticated backend connections. The deployment currently supports these private Agents:

  • Test Archive — 41 indexed files.

  • Bauer Kompressoren — 373 indexed files plus the synthetic Bauer Technical Twin search tool.

  • RapidDraft ERP + PLM — governed part, supplier, BOM, stock, purchase, lifecycle, document, and CAD-record search across the test Odoo and Aras datasets.

A user selects the Agent appropriate to the company, project, or system context before starting a conversation.

Open first

Page Primary question
ERP / PLM Integration Overview How are LibreChat, Odoo, Aras, CAD, RapidDraft, and Railway connected?
Data Ownership and Agentic Search Which system owns each record, and how must an Agent present results?
Operations and Testing How can the integration be tested safely?
Architecture and Boundaries Where does each service run?
Data and Storage Where are files, conversations, indexes, and vectors stored?
Knowledge Bases and Isolation How are companies separated and selected?
File Search and RAG Request Flow What happens during ingestion and during each question?
Models, Context, and Reliability Which models are used and where are the bottlenecks?
RAG Versions and V4 Direction What did each version prove, why did V3 underperform, and what will V4 change?
Bauer RAG V1 / V2 / V3 / V4 Benchmark What was actually measured, how does the repaired live V4 compare with preserved V1-V4 outputs, and what exact output was retained?
ONIX RAG Integration How can the same retrieval service be used from ONIX without copying LibreChat internals?
Operations Runbook How is the deployment checked, provisioned, and recovered?

Current system

Capability Current implementation State
User interface LibreChat at chat.rapiddraft.ai Deployed
Chat inference local/qwen-coder through LiteLLM Deployed
Embeddings local/embed-engineering, 1,024 dimensions Deployed
File retrieval Authorized one-request batch search through the LibreChat RAG API and Railway PostgreSQL/pgvector Deployed
RAG V2 retrieval Exact metadata, lexical, vector, table-aware, reranked, and evidence-validated retrieval in the existing RAG service Deployed for a private V2 Agent
RAG V3 retrieval Dedicated PostgreSQL evidence model and private API pinned to one ready candidate release Private development shadow; not promoted
RAG V4 Verified canonical evidence, simpler retrieval/reranking, full-question contract, and portable client adapters Planned; not implemented or deployed
Application records Railway MongoDB Deployed
Conversation search Railway Meilisearch Deployed
File objects Railway S3-compatible bucket librechat-files Deployed
Structured Bauer search Protected MCP/HTTP service, open business vocabulary, hard engineering constraints, and dedicated bauer_twin PostgreSQL database Deployed
Odoo operational records Live read through the Odoo JSON-2 API with database-scoped bearer authentication Deployed for the test dataset
Aras engineering records Versioned search snapshot derived from the Aras test dataset Deployed; private-network live read not yet enabled
ERP/PLM search Protected Railway gateway with HTTP and MCP interfaces, provenance, freshness, and system ownership Deployed
ERP/PLM conversational interface Private LibreChat Agent named RapidDraft ERP + PLM Deployed
CAD metadata and release planning Local workstation adapters and a browser Integration Lab; fixture mode and dry-run are the defaults Deployed for testing
RapidDraft DFM integration Authenticated model loading and live DFM service checks through the integration workflow Deployed for testing
Native RapidDraft embedded Agent Product-facing gateway backed by LibreChat Remote Agents Proposed; not enabled
Cloud inference fallback LiteLLM fallback to an EU cloud provider Proposed; not configured

Verified state

The LibreChat and knowledge-retrieval checks passed on 2026-07-22:

  • LibreChat, RAG API, and Bauer Twin health endpoints returned healthy responses.
  • All six Railway services reported successful deployments with one running replica.
  • Bauer Twin reported PostgreSQL mode with 12 projects, 75 parts, 8 documents, and 74 terminology aliases.
  • The Local AI chat, vision, embedding, LiteLLM, and Cloudflare Tunnel services were active.
  • Railway reported 417 file objects using 5.1 MB of S3 storage.
  • The MongoDB, Meilisearch, and PostgreSQL volumes were ready and persistent.
  • The pinned LibreChat and RAG overlays passed build-time upstream source checks and deployed successfully.
  • The batch-search acceptance run completed in 25.6 seconds with one RAG request per Agent query and disjoint file-ID sets.

The full ingestion acceptance run verified 41/41 Test Archive files, 373/373 Bauer files, zero shared file IDs, and grounded responses from both Agents. The Option B structured-search release passed 24 deterministic tests and 19/19 live Railway queries with zero hard-filter or status violations. Production Agent runs also returned the same nitrogen project for German and English queries and safely rejected an unavailable helium medium without weakening constraints.

The ERP/PLM integration checks passed on 2026-07-29:

  • The Odoo test dataset passed 19/19 content and relationship checks.
  • The Aras package contained the governed part, document, CAD, BOM, supplier, and lifecycle records required by the demonstration.
  • The search evaluation passed 8/8 representative questions and retained record IDs, source ownership, and freshness.
  • The Railway gateway passed 12/12 API tests, and both LibreChat MCP tools were discoverable.
  • A live LibreChat Agent run returned a grounded cross-system answer.
  • The local CAD bridge, live DFM service, browser Integration Lab, audit checks, and credential scan passed.

System boundaries

  • LibreChat is an operational knowledge interface, not the owner of RapidDraft product contracts.
  • Odoo owns operational product, supplier, stock, purchase, and manufacturing-planning records.
  • Aras owns engineering revisions, lifecycle state, documents, native CAD identity, and released engineering BOM intent.
  • Only released engineering revisions are eligible for publication to Odoo. Search does not bypass this gate.
  • The ERP/PLM search index is a rebuildable projection, not a new master-data system.
  • The Local AI Server owns model serving; it does not store LibreChat conversations or Railway vector data.
  • Railway owns the deployed application services and persistent LibreChat data.
  • The Bauer records are synthetic demonstration data and are not confirmed Bauer master data.
  • Knowledge-base isolation is logical. Independent legal, regulatory, administrative, or encryption boundaries require separate deployments.

Section map

Platform

Knowledge and models

Demonstration and operations

ERP / PLM integration

Product integration

Benchmarks

Open Questions

  • When should the Aras search projection move from a versioned snapshot to a private live adapter?
  • Which ERP/PLM roles and field-level restrictions are required before real company data is introduced?
  • Should the future native RapidDraft embedded Agent reuse the deployed ERP/PLM gateway?

Sources

  • D:\02_Code\LibreChat_Setup\README.md
  • D:\02_Code\LibreChat_Setup\librechat.yaml
  • D:\02_Code\15_ERP\docs\Agentic_Search_Integration_v1.md
  • D:\02_Code\15_ERP\docs\Phase_2_CAD_RapidDraft_Integration_Plan.md
  • D:\02_Code\15_ERP\docs\Implementation_Status_v1.md
  • Railway CLI status, service, volume, bucket, health, MCP, and Agent checks captured through 2026-07-27
  • Odoo, Aras, gateway, MCP, LibreChat Agent, CAD bridge, DFM, browser, audit, and credential checks captured on 2026-07-29
  • Local AI systemd status captured on 2026-07-21
  • V1/V2/V3 private-shadow benchmark captured on 2026-07-27