resilient-context-extraction

Resilient Context Extraction

Objective

Prevent data hallucination and inefficiency by mandating that agents inspect, validate, and fully extract data from provided reference files before attempting web searches or generating synthetic data. When extraction is incomplete, use fallback strategies before making assumptions.

Critical Rule

If a reference file is provided in the task context, it is the source of truth. Do not fabricate data or search the web for information that may exist within the provided attachments. If extraction appears incomplete, attempt alternative methods before proceeding.

Workflow Steps

1. Scan Context for Attachments

At the start of every task, explicitly list all files provided in the context window or attachment panel.

• Check for spreadsheets (.xlsx, .csv), documents (.pdf, .docx, .pptx), or data dumps (.json, .txt, .xml, .yaml).
• Note the filename, file size (if available), and inferred content type.
• Record this list for later verification.

2. Evaluate Relevance

Determine if any provided file contains the data required to complete the task.

• Match Keywords: Do filenames or expected column headers match task requirements?
• Check Scope: Does the data cover the necessary timeframe or region?
• Prioritize: Rank files by likelihood of containing required data.

3. Extract Data with Validation

If relevant files are found:

• Read the file content using appropriate tools (e.g., read_file, pandas, pdf_reader).
• Validate Extraction Completeness (NEW CRITICAL STEP):
  • Check if output appears truncated (e.g., sudden cutoff mid-sentence, character limits hit).
  • Compare expected data points vs. extracted data points (e.g., "Task mentions pricing tiers; did extraction include pricing numbers?").
  • Look for structural indicators of incompleteness (e.g., unclosed tables, missing document endings).
• If extraction is incomplete or suspicious, proceed to Step 4 (Fallback Strategies) before using the data.