Case Study · Hybrid AI Search

One Search. Every Answer.

Databases hold the figures. Documents hold the reasons. JKN Tech built an enterprise retrieval system that answers from both — exact numbers computed in your database, context cited from your own documents.

MyDB · Hybrid AI Research

Your Question

What was Q3 revenue growth for Product X, and what strategic factors caused the deviation from our projected budget?

route Intent Router → Hybrid Execution
database

Text-to-SQL · Exact Figures

Q3 revenue growth computed directly from the warehouse schema — no LLM arithmetic.

description

RAG · Cited Context

Deviation drivers retrieved from earnings transcripts and board decks — with source citations.

Synthesis Engine · Grounded Answer
Capabilities Text-to-SQL RAG & Semantic Search Context & Intent Routing Cross-Encoder Reranking Grounded Synthesis

The Problem

Your Data Lives in Two Silos

Modern enterprise data sits in two distinct worlds, each demanding a different analytical paradigm. Standard tools can only reach one of them.

table_chart

Structured Data

Relational databases and data warehouses. Transaction ledgers, inventory counts, time-series metrics.

check_circle

Exact numeric precision, schema enforcement, filtering and aggregation.

cancel

Blind to qualitative context and unindexed narrative text.

terminal

Reached by SQL queries.

description

Unstructured Data

PDF reports, emails, policy manuals, support tickets, call transcripts.

check_circle

Rich semantics, nuance, and the qualitative rationale behind decisions.

cancel

Poor at exact calculations, aggregations, and real-time figures.

manage_search

Reached by vector search and RAG.

The Insight Gap

"What was our Q3 revenue growth for Product X, and what strategic factors caused the deviation from our projected budget?"

Standard systems fail this question. A pure RAG pipeline struggles with exact financial calculation. A pure SQL pipeline returns the totals but cannot explain the qualitative why. The answer lives on both sides — and so does the retrieval.

The Solution

Hybrid AI Retrieval Architecture

A dual-engine framework that merges the deterministic precision of Text-to-SQL with the semantic flexibility of RAG — orchestrated by a Context & Intent Router.

Architecture diagram of the JKN Tech hybrid AI retrieval system: a natural-language query enters a Context and Intent Router, which dispatches to Text-to-SQL, RAG, or both in parallel, before a grounded synthesis engine produces the final answer.
The query path: one question in, routed to SQL, RAG, or both — synthesized into a grounded, attributed answer.
Engine A

database Text-to-SQL Pipeline

Converts natural language into validated SQL against your structured databases.

  • check Schema linking & dynamic table selection
  • check Dialect translation: PostgreSQL, Snowflake, BigQuery
  • check Sandboxed syntax & execution validation
  • check Read-only execution, column-level permission checks
Engine B

manage_search RAG Pipeline

Extracts narrative, contextual knowledge from your unstructured documents.

  • check Multi-stage chunking & metadata tagging
  • check Hybrid vector search: dense semantic + sparse BM25
  • check Cross-encoder reranking for relevancy scoring
  • check Contextual compression to trim token overhead
The Brain

route Context & Intent Router

Classifies every query and picks the optimal execution path — cutting latency and token cost instead of running everything through both pipelines.

  • calculate Quantitative intent → Text-to-SQL
  • notes Qualitative intent → RAG
  • sync_alt Composite intent → parallel hybrid execution
Deployed Product

Built for the People Asking the Questions

No query language. No hunting between BI dashboards and document folders. Executives, operations, sales, and compliance staff ask in plain language — and get figures with their reasons attached.

verified

Zero Math Hallucination

Numbers execute in the database engine via SQL. LLMs never compute totals.

receipt_long

Dual Attribution on Every Answer

The underlying SQL query and direct document citations travel with the response.

bolt

Reduced Search Latency

Simple queries skip unnecessary processing. The router sends each request down the shortest path.

Screenshot of the MyDB interface: a business user asks a plain-language question and receives a grounded answer drawing on both database metrics and cited document context.
The MyDB interface — one natural-language window over the entire corporate data footprint.

Why It Works

Hybrid Beats Either Engine Alone

Combining Text-to-SQL and RAG resolves the core limitation of using either technique in isolation.

Evaluation Dimension Pure Text-to-SQL Pure RAG JKN Tech Hybrid System
Numerical Accuracy Deterministic — accurate to database state Prone to LLM calculation hallucinations Deterministic SQL accuracy for all metrics
Qualitative Depth None — cannot read free text High — retrieves narrative context High — pairs numbers with narrative context
Query Flexibility Limited to predefined schemas Flexible across any written text Comprehensive across all corporate assets
Data Recency Real-time query execution Limited by indexing schedule Real-time database reads + indexed docs
Trust & Verification Transparent SQL query inspection Direct citation of source documents Dual attribution: SQL query + doc citations

Use Cases

Where the Hybrid Pays for Itself

The same architecture serves every role that has ever needed a number and the story behind it.

monitoring

Executive & Financial Reporting

"What was our EBITDA margin expansion in Q2, and what management commentary was provided regarding rising logistics costs?"

SQL pulls exact quarterly figures from the warehouse; RAG retrieves the reasoning from earnings-call transcripts and board decks.

local_shipping

Supply Chain & Operations

"How many units of Part #4821 are in stock across European warehouses, and what are the force majeure terms in our vendor contract with Supplier Y?"

Real-time inventory counts from the ERP, clause specifics retrieved from the vendor contract PDF.

support_agent

Customer Success & Contract Auditing

"What is Customer Z's current annual recurring revenue, and what are their custom SLA response commitments?"

Subscription details from the CRM, non-standard terms extracted from scanned agreement documents.

gavel

Regulatory & Compliance Auditing

"Which audit findings from last quarter remain unfulfilled, and what total financial penalties do they carry under current risk framework rules?"

Penalty thresholds from financial risk models, mapped to qualitative audit report summaries.

Enterprise Value

What the Hybrid System Delivers

block

Zero Math Hallucination

Every numerical query executes directly in your database engine. LLMs never compute totals, so calculation errors cannot enter your reporting.

speed

Reduced Search Latency

The intent router keeps simple database or document queries off unnecessary processing stages. Fast paths stay fast.

fact_check

Auditable Decisions

Every answer carries its evidence: the SQL query that produced the figures and footnoted citations for every source document.

hub

Unified Knowledge Access

One natural-language interface over BI and knowledge-management systems. The friction between the teams disappears.

Your Data Already Holds the Answers.

Put one interface in front of it. JKN Tech designs, deploys, and manages the hybrid retrieval stack on infrastructure you control — your data never leaves your perimeter.