Technical Appendix

Technical Appendix — AskVision

A. System Architecture Overview

AskVision is designed as a modular AI-first analytics system that translates natural language queries into structured market analysis.

High-level components

• Client Interface: Web UI, API, and optional chat-based clients

• Query Engine: Natural language parsing, intent detection, and parameter extraction

• Analysis Core: Indicator computation, pattern recognition, and inference logic

• Data Layer: Market data ingestion, normalization, and caching

• Execution & Output Layer: Result synthesis, formatting, and delivery

The architecture emphasizes stateless execution, deterministic analysis, and clear separation between AI interpretation and numerical computation.

---

B. Data Ingestion & Normalization

B.1 Data Sources

AskVision consumes:

• Spot and derivatives market data (OHLCV, funding, OI)

• Order book snapshots (where available)

• On-chain metrics (select networks)

• Derived indicators (precomputed and on-demand)

B.2 Normalization Pipeline

All incoming data is:

1. Timestamp-aligned

2. Converted to canonical units

3. Validated for completeness

4. Stored in a normalized internal schema

This ensures consistent outputs regardless of upstream data provider variance.

---

C. Query Interpretation Engine

C.1 Natural Language Parsing

User queries are parsed into:

• Asset scope (symbol, market, chain)

• Timeframe(s)

• Indicators or patterns

• Constraints (risk, volatility, trend bias)

Example internal representation:

QueryIntent {
  asset: BTC
  timeframe: 4H
  indicators: [RSI, VWAP]
  question_type: trend_confirmation
}

Ambiguous queries are resolved conservatively, prioritizing explicit user intent over inferred assumptions.

---

D. Analysis Core

D.1 Indicator Computation

Indicators are computed using:

• Fixed, documented formulas

• Explicit lookback windows

• Deterministic rounding rules

No indicator output is modified by AI post-processing.

D.2 Pattern & Signal Logic

Pattern recognition (e.g., trend structure, momentum divergence) is rule-based and auditable.

AI is used to:

• Select relevant analyses

• Explain results

• Highlight confidence and uncertainty

AI does not generate raw signals or fabricate data.

---

E. Execution Flow

1. User submits query

2. Query is parsed and validated

3. Required datasets are fetched or computed

4. Analysis engine executes deterministically

5. Results are passed to AI for explanation

6. Final response returned to user

Failures at any stage result in:

• No partial outputs

• Explicit error responses

• Logged diagnostics (non-user facing)

---

F. Security Model

F.1 Assumptions

• All external input is untrusted

• Market data providers may be inconsistent or delayed

• Users may attempt prompt injection or manipulation

F.2 Mitigations

• Strict query schema validation

• Sandboxed AI execution context

• Hard limits on compute and data scope

• No direct execution of user-supplied code

AskVision never executes trades, holds keys, or interacts with wallets.

---

G. Performance Characteristics

• Indicator computation: O(n) per timeframe window

• Query latency: Optimized via caching and precomputation

• Scalability: Horizontal scaling at query and analysis layers

High-load scenarios degrade gracefully by:

• Reducing explanation verbosity

• Preserving numerical accuracy

---

H. Error Handling & Transparency

Errors are structured and user-readable:

{
  "type": "DATA_UNAVAILABLE",
  "description": "Insufficient historical data for requested timeframe"
}

AskVision explicitly states when:

• Data is missing

• Confidence is low

• Results are inconclusive

---

I. Configuration & Deployment

• Environment-based configuration

• No hardcoded secrets or credentials

• Immutable builds recommended

Environment tiers

• Development: extended logging, experimental features

• Staging: production parity

• Production: rate limits, monitoring, hardened settings

---

J. Testing & Validation

J.1 Test Coverage

• Unit tests for indicators

• Regression tests for historical scenarios

• Property-based tests for invariants

• Adversarial prompt testing

J.2 Determinism Checks

Identical inputs must always produce identical numerical outputs, independent of AI explanation phrasing.

---

K. Upgrades & Versioning

• Semantic versioning enforced

• Indicator formula changes require major version bump

• Backward compatibility prioritized for APIs

Deprecated features are clearly marked and documented.

---

L. Known Limitations

• AskVision is not a trading bot

• Outputs are analytical, not predictive guarantees

• Market regime shifts may invalidate historical patterns

Users remain fully responsible for trading decisions.

---

M. Developer Resources

• Indicator formula references

• API schemas

• Example queries and outputs

• Validation datasets