Multi-Source Field Extraction Process
Step 1: Check Tool Availability
Confirm that the current session has both web_search and web_fetch tools. If either is missing, inform the user which tool is unavailable and stop the process.
Step 2: Search and Filter Sources
Call web_search using the user's original query. From the results, select 3~5 sources, prioritizing:
- Official sources (government ministry websites, corporate announcements, official statistics pages)
- Authoritative media (portal news, industry vertical media) Exclude clearly irrelevant ads, forum posts, and aggregator pages.
Step 3: Parallel Fetching
Call web_fetch for each selected URL. Issue all calls concurrently in a single response, not sequentially. If a URL fetch fails, log the failure reason and continue processing the remaining URLs.
Step 4: Extract Fields
From each successfully fetched page, extract the following fields:
- Date: The date directly related to the query topic (e.g., pricing change date, publication date). Do not extract footer copyright years.
- Source Name: The institution or media name to which the page belongs.
- Key Values/Facts: Specific numbers or conclusions directly related to the query topic (e.g., "increased by XX yuan/ton"). If a page lacks a certain field, mark it as "not found" — do not fabricate.
Step 5: Consolidate Output
Output in a table format, one row per source:
| Source | Date | Key Values/Facts | Fetch Status |
|---|
Append after the table: Consensus Information: If multiple sources agree on the same fact, list it. Conflicting Information: If values or dates differ across sources, list them and annotate each with its source.
Boundaries
- Do not summarize the full page content; only extract fields relevant to the query topic.
- If all sources fail to fetch, directly report the failure; do not use search result snippets as substitutes for page content.
- Do not recommend which source the user should consult; only perform information consolidation.