Portfolio
AI Systems

Query

Private RAG for search, selection, and execution

May 18, 2026exec · dev · ai-curious
Query screenshot 1: query5

Outcome

Match user intent to the right model, workflow, or answer path while keeping institutional knowledge private

Innovation

Private RAG with context-aware model and action selection across tools and workflows

Stack

LLMs · Retrieval · Automation · Evaluation · Prompting
Query screenshot 2: query1
Query screenshot 3: query2
Query screenshot 4: query3
Query screenshot 5: query4
Query screenshot 6: query mobile

Query is designed for the messy middle between a user request and the actual system behavior needed to satisfy it.

Instead of treating every prompt as the same kind of problem, Query classifies intent, evaluates context, and selects the right answer path.

The Challenge

  • Different questions need different systems
  • Users rarely describe the task in machine-ready terms
  • Single-model approaches flatten important distinctions
  • Tool invocation and model choice often remain manual

The Query Model

Query treats an incoming request as a routing problem.

  1. Interpret the request and surrounding context
  2. Classify the underlying task shape
  3. Select the right model, tool, or workflow path
  4. Execute with constraints and observability
  5. Learn from outcomes and correction patterns

Technical Architecture

Private RAG Flow

Query logical architecture

Request LayerUser intent + context

Captures the question, surrounding context, and task shape before selecting an answer path.

Interpretation
Intent classificationContext shapingTask routing
Decision CorePrivate retrieval + model / tool selection

Retrieves internal knowledge, selects the right model or workflow, and keeps institutional memory out of public model state.

Governance
Access boundariesGrounded answersExecution constraints
Knowledge Layer
Private documentsChunk / citation storeInstitutional memory
Delivery Channels
ChatSMSVoiceWorkflow handoff
Learning Loop
Feedback captureOutcome comparisonSelection refinement
  • Intent-first routing: each request is interpreted before the system chooses a model, tool, or workflow.
  • Private retrieval layer: internal knowledge stays grounded in controlled document context instead of leaking into public model memory.
  • Multi-channel delivery: the same decision layer can serve chat, SMS, voice, and operational handoff paths.
  • Learning loop: feedback and outcome patterns refine future selection behavior over time.

Core Capabilities

1. Intent Understanding

Query identifies whether the request is explanatory, analytic, generative, operational, or automatable.

  • Semantic task understanding
  • Context-aware interpretation
  • Request shaping for downstream systems

2. Model and Tool Selection

The system chooses the best path instead of assuming one universal model is enough.

  • Cost-aware model selection
  • Capability-aware routing
  • Tool and workflow handoff
  • Policy-aware execution boundaries

3. Private Retrieval

Query can use retrieval-augmented generation to keep institutional knowledge usable without pushing it into public or shared model memory.

  • Private knowledge retrieval
  • Grounded answers from internal sources
  • Controlled access to sensitive context
  • Institutional memory without public exposure

4. Operational Learning

Query improves over time by studying what worked.

  • Feedback capture
  • Outcome comparison
  • Repeated-pattern detection
  • Selection refinement over time

Why It Matters

The value of AI often depends less on any single model than on selecting the right path for a given request.

Query is built around that selection layer.