From Unstructured Text to
Verified Reality

A technical walkthrough of the Agentic RFP Analysis architecture. Moving beyond simple summarization, this system atomizes requirements, enforces strict data hygiene, and creates an audit-proof "Evidence Interlock" for high-stakes enterprise bidding.

The 7 Pillars of Truth

The architecture implements a seven-stage pipeline to convert unstructured RFP documents into structured data. It establishes clear lineage from raw requirements to validated capabilities, ensuring traceability and evidence-based verification throughout the process.

STATUS: ACTIVE

Demand Signal

Ingestion & Atomization

Atomizes unstructured text into immutable "Requirement Atoms" with verbatim anchors.

Ingests raw RFP documents (PDF, DOCX) and uses NLP to segment text into discrete, traceable atomic units.

System Metric100% Traceability
ID: 01_SYS
STEP 1: DATA STRUCTURE

Atomization Logic

Before semantic understanding, we must achieve Structural Integrity. The system employs a "Docling" layout parser to capture metadata context (page numbers, section headers) before breaking content into atoms.

Why this matters: A standard LLM sees a "soup of text." By parsing the layout first, we know that "10 days" isn't just a numberβ€”it's a constraint belonging to Section 4.1.

Each atom keeps a verbatim anchor to the legal source while being stored in an append-only ledger. Amendments create new atoms that supersede prior versions, Q&A items clarify active requirements, and any unmatched references trigger an orphan check for human review.

Raw Input (RFP Page 42)

"The Contractor shall provide a dedicated Project Manager who possesses a current PMP certification... and who must be available to report on-site within ten (10) days of contract award."

RFP Document - Page 42
"The Contractor shall provide a dedicated Project Manager who possesses a current PMP certification... and who must be available to report on-site within ten (10) days of contract award."
Explore the Ingestion Pipeline
Impact of Time on Trust Scores
Figure 2: The Time-Decay Impact. Note how a "Gold Contract" > 3 years old scores lower (55) than a fresh Silver proposal (75), enforcing content freshness.
STEP 2: HYGIENE PROTOCOL

The Garbage-In Problem

Not all data is equal. The Hygiene Protocol calculates a composite Quality Vector based on three factors: Source Verification, Time Decay (Freshness), and Usage Frequency.

Gold Tier (Verified) High Trust

Signed contracts and audited financials.
*Subject to time decay.

Silver Tier (Probable) Med Trust

Submitted proposals and resumes < 6 months old.

Bronze Tier (Risk) Low Trust

Marketing slicks, wiki drafts, and outdated assets.

STEP 3: MATCHING LOGIC

Two-Pass Reconciliation

Finding the right evidence requires balancing speed and accuracy. We separate retrieval from adjudication using a Vector Triangulation phase followed by a Critic Agent review.

The Filtering Funnel

Reconciliation Funnel

Phase A: Vector (The Net)

Rapidly scans thousands of assets using Cosine Similarity.

"Find anything mentioning Oracle and Postgres."

Phase B: Critic (The Filter)

Applies deep reasoning to check the Verbatim Anchor.

"Discard results that don't explicitly show migration FROM Oracle TO Postgres."

STEP 4: TRACEABILITY

The Evidence Interlock

Complex requirements rarely map to a single document. The system creates a Composite Proof (or "Zipper") that clusters multiple validated assets.

Interactive Demo: Select a scenario to see how the system clusters evidence or flags failure when an interlock is missing.

Scenario: Key Personnel
Req: Quality Manager (ISO 9001 + 10 Yrs Exp)
πŸ‘€
Resume: Jane Doe (20 Yrs Exp)
πŸ“œ
Cert: ISO 9001 Lead Auditor (2023)
🏒
Org Chart: Role Assignment
STATUS: VERIFIED

Asset Selection Logic

X: Freshness (Days) | Y: Relevance Score | Color: Trust Tier

Loading Visualization...
The system prioritizes the "Top Right" quadrant: High Relevance + Recent Freshness.
STEP 5: VALIDATION

Operational Gateways

Automation is not abdication. The system implements hard-coded Human-in-the-Loop checkpoints. This is not just a "review"β€”it's a control plane where human rejection forces the AI to loop back and retry.

Gateways also resolve orphaned amendments and approve supersedes/clarifies lineage links before any requirement is treated as binding.

Interactive Demo: See what the human operator actually controls at each stage of the lifecycle.

Operator Dashboard View
Gateway 1: Structure Review
Pending Approval
Original Text Segment:

"Contractor must be US Citizen and possess TS/SCI."

Atomization Result:
βœ“Possess TS/SCI
!US CitizenMissing Atom?