Documented Build · v1.0
RFP Response & Sales Evidence Control Room
Turns a long RFP into a requirement matrix, retrieves only approved company evidence, drafts citation-backed answers, and routes risky commitments to a human.
Documented RAG build tested with fictional evidence on n8n, NVIDIA embeddings and chat, and a local Qdrant collection. No customer or win-rate claim is made.
DOC-SEC-001Security controls · v3.2
0.94DOC-SLA-001Service levels · v2.0
0.42Customer data is encrypted in transit with TLS 1.2 or later and at rest using AES-256. [DOC-SEC-001]
The controlled RAG path
Every answer must survive the evidence gate.
Select a stage to see what the system is allowed to do. Knowledge permission and proposal approval remain outside the model.
Register
Documents are validated for approval state, expiry, classification, region, version, and checksum before NVIDIA creates embeddings.
Extract
The RFP becomes a requirement matrix that preserves obligations, numbers, category, mandatory status, risk hints, and retrieval queries.
Retrieve
Qdrant returns semantically relevant evidence while status, expiry, classification, and project permissions are enforced before drafting.
Control
Deterministic code checks citations, unsupported claims, confidence, category, and sensitive commitment language after the model responds.
Assemble
The response package includes only approved answers, blocked requirement IDs, and a deduplicated evidence manifest for final human review.
Policy outside the model
High confidence is not permission.
The drafting model reports confidence and evidence IDs. Deterministic code then decides whether the answer can proceed or must enter review.
Operational value
A reusable answer system—not a faster copy machine.
The design improves speed without removing evidence ownership, SME judgment, or a final release decision.
Controlled evidence
Only approved, non-expired documents with classification and region metadata can enter the response path.
Traceable answers
Every draft carries evidence IDs, confidence, unsupported-claim flags, and deterministic control reasons.
Human commitments
Pricing, legal terms, guarantees, low confidence, and missing evidence route to an owner or SME before assembly.
Owner and SME review
Resolve risky language from Telegram.
Only escalated answers reach the reviewer. This visual simulator changes the example state; it does not send a message.
Evidence: none
Confidence: 0.32
Reasons: NO_APPROVED_EVIDENCE, SENSITIVE_COMMITMENT
Accept the grounded draft
The answer becomes eligible for proposal assembly with its evidence IDs and reviewer decision preserved.
Approved answer + audit eventEnterprise boundary
The trust model is visible and enforceable.
The starter is intentionally simple enough to understand, while preserving the controls an enterprise implementation needs to extend.
- The language model cannot approve its own answer or submit a proposal.
- Expired, unapproved, or disallowed-classification evidence is removed before drafting.
- Pricing, legal terms, guarantees, warranties, and similar commitments always require review.
- Public templates contain no credentials, live workbook IDs, owner identities, or customer documents.
- Multi-tenant or highly sensitive deployments should use separate collections or deployments—not prompts as access control.
Free starter files
Download the whole evidence-control system.
Every public workflow is inactive, credential-free, and checked before packaging. The evidence library and RFP are fictional.
Complete starter bundle
Six sanitized workflows, Qdrant compose file, operations workbook, fictional evidence, fixtures, security guide, and QA report.
Download ZIPOperations workbook
Nine controlled tabs for projects, requirements, evidence, reviews, exceptions, audit history, approved answers, and metrics.
Download XLSXProduction error handler
Masked Telegram/email alerts and exception logging.
Download JSONKnowledge library manager
Approved-source gate, one-document loop, NVIDIA embeddings, Qdrant ingestion, and optional source adapters.
Download JSONRFP intake and extractor
Manual fixture and webhook intake, validation, NVIDIA extraction, and requirement-matrix mapping.
Download JSONRetrieval and drafting engine
Qdrant retrieval, permission filtering, evidence-bound drafting, and deterministic review policy.
Download JSONApproval and SME dispatcher
Telegram/email review cards plus approve, reject, and request-edit state handling.
Download JSONProposal assembly
Approved-answer assembly, blocked-item report, evidence manifest, and disconnected publishing adapters.
Download JSONQdrant Docker service
Pinned Qdrant service attached to the existing n8n Docker network.
Download YAMLSanitized release: no NVIDIA key, Qdrant secret, Google account, workbook ID, Telegram identity, SMTP credential, customer RFP, or private knowledge document is included.
Implementation manual
From empty vector store to reviewed proposal package.
RFP Evidence Control Room
This is a governed RAG starter for proposal teams. It registers approved knowledge, extracts atomic RFP requirements, retrieves permitted evidence, drafts cited answers, routes risky language to a human, and assembles only approved responses.
The six modules
- Production error handler
- Knowledge library manager
- RFP intake and requirement extractor
- Evidence retrieval and drafting engine
- Risk, approval and SME dispatcher
- Proposal assembly and operations
The modules share project IDs, requirement IDs, the operations workbook, and Qdrant collection. They are separate so each trigger, permission boundary, retry policy, and scaling path stays visible.
All bundled documents, customers, commitments, URLs, and proposal content are fictional.
What to download
Start with rfp-sales-evidence-control-room-v1.0.zip. It contains every workflow, the workbook, Qdrant compose file, fictional fixtures, security guidance, and the full implementation manual.
Use the individual JSON downloads when you only need to replace one workflow.
Important
- Public workflows are inactive.
- Credential bindings are removed.
SET_*values must be replaced.- The expired pricing fixture is intentionally included to test rejection.
- The system never submits a proposal automatically.
Installation
- Start n8n 2.30.5 or later.
- Attach the Qdrant service to the same Docker network as n8n.
- Import
rfp-control-room-operations.xlsxinto Google Sheets. - Import the six workflow JSON files in numeric order.
- Create NVIDIA, Qdrant, Google Sheets, Telegram, and SMTP credentials.
- Replace every
SET_*placeholder. - Keep all workflows inactive until the test tab passes.
Run the Knowledge Library Manager first. The Retrieval and Drafting Engine requires an indexed collection.
For the exact Docker checks, node bindings, workbook tabs, and activation order, use the included IMPLEMENTATION-GUIDE.md.
Credentials
NVIDIA
Bind the same NVIDIA credential to both chat nodes and both embedding nodes.
- Chat:
nvidia/llama-3.3-nemotron-super-49b-v1.5 - Embeddings:
nvidia/nv-embedqa-e5-v5
Qdrant
For the included Docker service, use http://rfp-qdrant:6333 from n8n. Do not expose an unauthenticated Qdrant port to the internet.
Google, Telegram and SMTP
Google Sheets needs edit access to the operations workbook. Telegram needs the owner chat ID and a stable HTTPS webhook for local n8n. SMTP sends only owner alerts in the starter.
Changing the embedding model requires a new collection and full re-index.
Knowledge library
Every source must map to the canonical control fields: document ID, title, department, classification, region, status, version, valid-until date, source URL, checksum, and content.
Only APPROVED, non-expired documents enter Qdrant. Each document is processed in its own loop iteration so n8n AI sub-node metadata resolves against the correct item.
Connect Google Drive, SharePoint, Confluence, S3, or a database through a source-specific mapper. Preserve upstream permissions and never widen classification during mapping.
Production installations also need withdrawal/tombstone handling and scheduled reconciliation between the registry and Qdrant.
RAG controls
The model receives only evidence that passes status, expiry, classification, and project-permission checks.
After drafting, deterministic code checks:
- Evidence count
- Citation presence
- Citation validity
- Unsupported claims
- Confidence threshold
- Legal or commercial category
- Pricing, discount, guarantee, indemnity, liability, penalty, warranty, zero-incident, and unlimited-commitment language
The default auto-approval threshold is 0.78, but model confidence is not a calibrated probability. Tune it with a labelled evaluation set.
The model cannot override these rules or approve itself.
Approvals
The dispatcher sends only escalated answers to Telegram and owner email.
The review card includes the requirement, draft, evidence IDs, confidence, unsupported-claim flags, and deterministic control reasons.
- Approve makes the answer eligible for assembly.
- Request edit keeps the requirement blocked.
- Reject excludes the answer.
Before production, add reviewer allowlists, approval expiry, stale-action rejection, idempotency, fresh row reads, named SME routing, and separation of duties for legal and commercial decisions.
Test the system
Use only the fictional fixture first.
- Confirm three approved documents are embedded.
- Confirm the expired pricing document is excluded.
- Extract all five sample requirements.
- Retrieve
DOC-SEC-001for the security question. - Confirm a grounded cited answer can proceed.
- Confirm pricing and zero-incident guarantees route to review.
- Test approve, reject, request edit, duplicate callbacks, and stale callbacks.
- Confirm assembly excludes pending and rejected answers.
- Force a provider error and verify the safe alert.
The downloadable QA-REPORT.md records automated and live checks.
Operate and recover
Select workflow 00 as the error workflow for modules 01–05.
Recommended defaults:
- Ten-minute timeout
- Execution progress enabled
- Error executions retained
- Successful execution retention limited to operational need
- Ninety-day maximum execution retention unless policy requires less
- Concurrency matched to NVIDIA quotas
Back up n8n, the Google workbook, and the Qdrant volume. Test restore. Re-index into a new collection after changing the embedding model.
Monitor open exceptions, pending reviews, expired documents, missing owners, retrieval failures, and provider latency.
Enterprise adaptation
For multi-tenant or highly sensitive use, prefer separate collections or deployments per security boundary. Prompt instructions are not access control.
Add:
- SSO, MFA, least privilege, and external secrets
- Private networking, TLS, Qdrant authentication, and encrypted backups
- Signed intake webhooks, rate limits, size limits, malware scanning, and file allowlists
- Dev/test/prod separation and change control
- Queue workers, transactional state, central logging, and disaster recovery
- Data residency, retention, deletion, legal hold, and model-provider assessment
- Labelled retrieval and answer evaluations with rollback
Keep customer submission as a separate, authorized step after final proposal review.