White Paper 5: High-Fidelity Search & Reflective RAG

1. Semantic Search Restoration

Traditional search engines often fail because they lack "Semantic Context." see7 restores this via our Hybrid Search Model.

Beyond Keyword ILIKE: While we maintain a PostgreSQL ILIKE fallback for emergency reliability, our primary engine uses pgvector Cosine Similarity. This allows see7 to understand the intent of a query (e.g., "Show me our pricing for growth-stage companies") even if the exact words "pricing" and "growth" are not present in the same sentence.

The Normalization Gate: Raw vector distances are abstract and difficult for humans to trust. see7 normalizes these into a 0–100% scale. Any snippet falling below our Sovereign Noise Floor (typically 0.4) is discarded, ensuring the AI only synthesizes answers from high-confidence data points.

2. The Reflective RAG Protocol

Synthesis without verification is dangerous. see7 implements Reflective RAG, a multi-stage process that prioritizes "Truth" over "Fluency."

Stage 1: Verified Snippet Prioritization (The Librarian's Domain): The engine does not treat all data equally. It first scans for Snippets marked as isGolden. These Golden Truths are created and curated by the LIBRARIAN role. Through the see7 Knowledge Vault, Librarians "bless" specific facts—official pricing, verified technical specs, or approved legal clauses. These verified facts are given a 2× weight in the synthesis prompt, ensuring the AI "anchors" its answer to the official record.

Stage 2: The Deep Reflection Loop: Before an answer reaches the user, the AI performs a "Self-Audit" using a reasoning-heavy model (Gemini 2.0 Pro). The system generates a draft, then cross-references every claim against the retrieved source snippets.

If a claim (e.g., "The product costs $500") is not explicitly supported by the retrieved snippets, the loop strikes the claim and re-drafts.

This process repeats until the response achieves a Verification Score of >0.9. If it cannot achieve this, it informs the user: "I found relevant information, but I cannot verify the specific figure you requested," maintaining our commitment to Standard #7 (Enterprise Grade).

3. High-Fidelity Traceability (Standard #6)

see7 maintains a 1:1 "Source-to-Claim" mapping that is unique in the AIaaS market.

Forensic Citation IDs: Every paragraph generated by the Drafting Desk (see White Paper 6) or Unified Search is tagged with forensic breadcrumbs. Users see active citation numbers that link directly to the specific Snippet record in the database.

[RAG:REFLECT] Logging: Every synthesis event is logged with a correlation ID and a [RAG:REFLECT] prefix. This log includes the "Quality Score" assigned by the reflection loop, providing administrators with a forensic audit trail of how the AI arrived at its conclusion.

4. Tunable Relevance Thresholds

Relevance is not "one-size-fits-all." In see7, the "Search Sensitivity" is a multi-layered configuration.

Global vs. Local Alignment: On a global scale, see7 enforces a baseline "Noise Floor." However, we are engineering a future feature where ADMINISTRATOR role users can tune these thresholds at a Customer Scale.

Role-Based Flexibility: This will allow a legal-heavy tenant to set a "High-Certainty" threshold (0.85+), while a marketing-heavy tenant might lower the threshold (0.60) to allow for more creative synthesis during ABM prospecting.

5. The "Truth-Anchored" User Experience

We replace the "Chatbot" feel with a Professional Grade Workbench.

Relevance Badges: Results are displayed with color-coded Relevance Badges (Green >80%, Yellow >60%, Gray ≤60%). This allows a sales rep to immediately know if an answer is a "Slam Dunk" or if it requires more human verification.

Drafting Tray Integration: Verified search results can be "pinned" to a drafting tray. This acts as a bridge to the Drafting Desk, where users can build complex RFPs manually from a library of verified facts, ensuring the human remains the final arbiter of high-stakes output.

Related White Papers

For ingestion and the Atomizer, see Omnivore Ingestion Engine. For the Drafting Desk (White Paper 6), see Drafting Desk Apps (coming soon). For engineering standards, see Development Philosophy.