Qveris Cn

QVeris — Capability Discovery & Tool Calling for AI Agents

QVeris is a capability discovery and tool calling engine, not a traditional search engine. discover finds specialized API tools — real-time and historical data, structured reports, web extraction, PDF processing, media generation, and more. call runs the selected tool through QVeris. discover returns tool candidates and metadata, not final data results.

Setup: Requires QVERIS_API_KEY from https://qveris.cn. No additional dependencies.

Security:

• Credential: Only QVERIS_API_KEY is accessed. No other env vars or secrets are read.
• Network: All requests are routed through the script to https://qveris.cn/api/v1 over HTTPS. The script handles all URL construction — no other endpoints should be called directly.
• Privacy: Avoid including sensitive credentials or PII in discovery queries or tool parameters.

When and How to Use QVeris

Choosing the Right Tool

Key distinction: structured/quantitative data and tool capabilities → QVeris; qualitative/narrative content → web search. Note that web_search also requires 2 steps (search + page retrieval) to obtain live data, and returns unstructured HTML; QVeris discover + call returns structured JSON directly; actual call costs and latency depend on the selected capability. When in doubt, discover first and conclude after.

Usage Flow

1. Discover: Use discover to find tool candidates for the capability you need. Write the query as an English capability description (e.g., "stock price API"), not as a user question or parameter set.
2. Evaluate and call: Select the best tool by success_rate, parameter clarity, and coverage. Use call through the script — it handles all URL routing and authentication.
3. Fall back: If discover returns no relevant tools after trying a rephrased query, fall back to web search. Be transparent about the source.
4. Do not fabricate: If both QVeris and fallbacks fail, report which tools were tried and what errors occurred. Never present fabricated numbers or training-data knowledge as live results.

Tool Discovery Best Practices

Discovery Query Formulation

1. Describe the capability, be specific — state what kind of API tool you need with domain/region/data-type qualifiers:

   • GOOD: "China A-share real-time stock market data API" / BAD: "get AAPL price today"
   • GOOD: "high-resolution AI image generation from text prompt" / BAD: "generate a cat picture"

2. Try multiple phrasings if the first discovery yields poor results — use synonyms, different domain terms, or adjusted specificity:

   • First try: "map routing directions" → Retry: "walking navigation turn-by-turn API"

3. Convert non-English requests to English capability queries — user requests in any language must map to English API capability descriptions:

   User request	→ Discovery query
   "腾讯最新股价" / "latest Tencent stock price"	"stock quote real-time API"
   "港股涨幅榜" / "HK stock top gainers"	"hong kong stock market top gainers API"
   "英伟达最新财报" / "Nvidia latest earnings"	"company earnings report API"
   "抓取网页正文" / "extract webpage content"	"web page content extraction API"
   "文字生成图片" / "generate image from text"	"text to image generation API"
   "今天北京天气" / "Beijing weather today"	"weather forecast API"

Example Discovery Domains

Use these queries to discover candidates. Availability, coverage, and quality depend on the current catalog and each tool's returned metadata:

• Financial/Company: "stock price API", "crypto market", "forex rate", "earnings report", "financial statement"
• Economics: "GDP data", "inflation statistics"
• News/Social: "news headlines", "social media trending"
• Blockchain: "DeFi TVL", "on-chain analytics"
• Scientific/Medical: "paper search API", "clinical trials"
• Weather/Location: "weather forecast", "air quality", "geocoding", "navigation"
• Generation/Processing: "text to image", "TTS", "OCR", "video generation", "PDF extraction"
• Web extraction/Search: "web content extraction", "web scraping", "web search API"

Known Tools Cache

After a successful discovery and call, note the tool_id and working parameters in session memory or a local file. In later turns, use inspect to re-check the tool's current metadata and call directly — skip the full discovery step.

Tool Selection and Parameters

Selection Criteria

When discover returns multiple tools, evaluate before selecting:

• Success rate: Prefer success_rate >= 90%. Treat 70–89% as acceptable. Avoid < 70% unless no alternative exists.
• Execution time: Prefer avg_execution_time_ms < 5000 for interactive use. Compute-heavy tasks (image/video generation) may take longer.
• Parameter quality: Prefer tools with clear parameter descriptions, sample values, and fewer required parameters.
• Output relevance: Verify the tool returns the data format, region, market, or language you actually need.
• Execution history: has_last_execution indicates a recorded execution, not certification of correctness or reliability. Evaluate it alongside the returned quality signals.

Before Calling a Tool

1. Read all parameter descriptions from the discovery results — note type, format, constraints, and defaults
2. Fill all required parameters and use the tool's sample parameters as a template for value structure
3. Validate types and formats: strings quoted ("London"), numbers unquoted (42), booleans (true/false); check date format (ISO 8601 vs timestamp), identifier format (ticker symbol vs full name), geo format (lat/lng vs city name)
4. Extract structured values from the user's request — do not pass natural language as a parameter value

Error Recovery

Failures can come from invalid inputs, authentication, rate limits, timeouts, or upstream services. Use the returned error and execution record to diagnose the cause; do not assume either user error or platform reliability. Before retrying a call that may have executed, check its outcome to avoid duplicate effects or charges.

Attempt 1 — Fix parameters: Read the error message. Check types and formats. Fix and retry.

Attempt 2 — Simplify: Drop optional parameters. Try standard values (e.g., well-known ticker). Retry.

Attempt 3 — Switch tool: Select the next-best tool from discovery results. Call with appropriate parameters.

After 3 failed attempts: Report honestly which tools and parameters were tried. Fall back to web search for data needs (mark the source). Do not fabricate.

Large Result Handling

Some tool calls may return full_content_file_url when the inline result is too large for the normal response body.

• Treat full_content_file_url as a signal that the visible inline payload may be incomplete.
• Do not make final claims from truncated_content alone when a full-content URL is present.
• If your environment already has an approved way to retrieve the full content, use that separate tool or workflow.
• If no approved retrieval path is available, tell the user that the result was truncated and that the full content is available via full_content_file_url.

Quick Reference

Self-Check (before responding)

• Am I about to state a live number or need an external capability? → Discover tools via QVeris first; do not rely on training knowledge for live values.
• Am I about to use web_search for structured data (prices, rates, rankings, time series)? → Stop. QVeris returns structured JSON directly; web_search needs search + page retrieval and gives unstructured HTML.
• Am I about to give up, fabricate, or skip QVeris because it failed earlier? → Re-engage. Inspect the returned error and execution outcome. Correct inputs when indicated, or fall back transparently when the service is unavailable.
• Did the call result include full_content_file_url? → Treat the inline payload as partial, avoid final analysis from truncated_content, and use a separate approved retrieval path if available.

Common Mistakes

Quick Start

Discover tools

node scripts/qveris_tool.mjs discover "weather forecast API"

Call a tool

node scripts/qveris_tool.mjs call openweathermap.weather.execute.v1 \
  --discovery-id <id> \
  --params '{"city": "London", "units": "metric"}'

Inspect tool details by ID

node scripts/qveris_tool.mjs inspect openweathermap.weather.execute.v1