Multi-Search Workflow

Run several searches with different phrasings to maximise recall, then merge and deduplicate before analysis.

Why Multiple Searches?

Semantic search matches conceptually similar content, but different phrasings cast different nets. A single search for “budget concerns” might miss documents about “cost overruns” or “financial pressure”. Running multiple searches with varied phrasing catches more relevant results.

Step-by-Step

1. Create an Evidence Directory

mkdir -p evidence

2. Run Multiple Semantic Searches

Use different phrasings for the same concept:

uv run python query.py \
  --semantic "project delays timeline concerns" \
  --top-k 200 --export-json evidence/delays.json
 
uv run python query.py \
  --semantic "budget overrun cost escalation" \
  --top-k 200 --export-json evidence/budget.json
 
uv run python query.py \
  --semantic "stakeholder feedback complaints" \
  --top-k 200 --export-json evidence/feedback.json

3. Combine Metadata Filters

Add sender, date, or source-type filters to narrow results:

uv run python query.py \
  --semantic "project risk" \
  --sender "[email protected]" \
  --date-range 2025-01-01 2025-06-30 \
  --top-k 100 --export-json evidence/pm-risks.json

4. Merge and Deduplicate

The merge tool removes duplicate documents found by multiple searches:

uv run python merge.py evidence/*.json --output evidence/merged.json

5. Triage Separately

Run triage as a standalone step so results are saved to disk. This lets you re-run deep analysis at different thresholds without re-triaging:

# Recommended: gemini-flash for triage (cheapest, fastest)
uv run python analyze.py evidence/merged.json \
  --triage \
  --model gemini-flash \
  --truncate 500 \
  --concurrency 5 \
  --context "Key project risks and stakeholder concerns" \
  --output evidence/triaged.json \
  --dry-run
 
# Free, private, slow (local Mistral 7B):
uv run python analyze.py evidence/merged.json \
  --triage \
  --local \
  --context "Key project risks and stakeholder concerns" \
  --output evidence/triaged.json

Use gemini-flash for triage (cheapest, fastest). Use --truncate 500 and --concurrency 5 for speed. Checkpoints save every wave — if interrupted, re-run the same command to resume.

6. Deep Analysis (on triaged results)

Use --deep-only to skip triage and run deep analysis on already-triaged data:

uv run python analyze.py evidence/triaged.json \
  --deep-only \
  --min-relevance 5 \
  --context "Key project risks and stakeholder concerns" \
  --model deepseek \
  --dry-run

Review the cost estimate, then run without --dry-run. If the output is too noisy, re-run at --min-relevance 7 (no re-triage needed).

Best Practices

• Start broad, then narrow — cast a wide net first, then filter
• Use 3-5 different phrasings per topic for good coverage
• --top-k 200 is a good default — adjust based on corpus size
• Always triage separately with --triage --output so results are saved
• Start with --min-relevance 5 to avoid missing borderline evidence
• Always --dry-run before paid analysis to check costs
• Use gemini-flash for triage (cheapest, fastest), deepseek for deep analysis (best reasoning)
• Use --truncate 500 and --concurrency 5 for fast triage
• Use --retry-failed to re-triage failed batches without re-running everything
• Triage checkpoints save progress — re-run to resume if interrupted