kuavi-deep-search

name	kuavi-deep-search
description	Multi-pass search with query refinement for hard-to-find video content. Use when initial search returns low-confidence results or no matches.

Deep Search with Query Refinement

When a simple kuavi_search_video call returns poor results (scores < 0.3 or no relevant matches), escalate through these refinement passes.

Pass 1: Field Rotation

Try different search fields — the same query may match differently depending on how content was indexed:

kuavi_search_video("your query", field="summary", top_k=5)
kuavi_search_video("your query", field="action", top_k=5)
kuavi_search_video("your query", field="visual", top_k=5)
kuavi_search_video("your query", field="temporal", top_k=5)

If field="summary" misses, field="visual" may catch it through raw frame embeddings.

Pass 2: Query Reformulation

Rephrase the query with synonyms, more general terms, or more specific terms:

Original	Reformulations
"person cooking pasta"	"making food", "kitchen activity", "preparing meal"
"final score"	"scoreboard", "results table", "end of game"
"presenter introduction"	"title slide", "speaker name", "opening remarks"

Try 2-3 reformulations across different fields.

Pass 3: Transcript Search

Switch from visual to audio modality:

kuavi_search_transcript("keyword from question")

Transcript search is keyword-based — use specific nouns, names, or distinctive terms rather than descriptions.

Pass 4: Hierarchy Level Search

Search at coarser temporal granularity for broad localization:

kuavi_search_video("query", field="all", level=1, top_k=5)

Level 1 segments are ~30s chunks. Once you find the right region, search again at level 0 within that range.

Pass 5: Exhaustive Scan via kuavi_eval

As a last resort, compute similarity against ALL segments programmatically:

kuavi_eval("""
scenes = get_scene_list()
best_score = -1
best_scene = None
for s in scenes:
    # Check caption content directly
    caption = s.get('caption', '') + ' ' + str(s.get('annotation', ''))
    if 'keyword' in caption.lower():
        print(f"Direct match: scene {s['scene_index']} at {s['start_time']:.1f}s")
        print(f"  Caption: {caption[:200]}")
""")

Pass 6: Temporal Elimination

If you know when the content is NOT (e.g., "not in the first half"), search only the remaining region:

kuavi_eval("""
# Search only the second half of the video
scenes = get_scene_list()
mid = scenes[-1]['end_time'] / 2
late_scenes = [s for s in scenes if s['start_time'] > mid]
for s in late_scenes:
    print(f"Scene {s['scene_index']}: {s['start_time']:.1f}-{s['end_time']:.1f}s: {s.get('caption', '')[:100]}")
""")

Decision Flow

Initial search (field="all") → scores > 0.4? → DONE
                             ↓ no
Field rotation (summary/action/visual/temporal) → found? → DONE
                             ↓ no
Query reformulation (3 variants) → found? → DONE
                             ↓ no
Transcript search → found? → DONE
                             ↓ no
Hierarchy level 1 search → found region? → search level 0 within region → DONE
                             ↓ no
Exhaustive scan via kuavi_eval → found? → DONE
                             ↓ no
Report: "Content not found after exhaustive search across all fields and modalities."

Key Principle

Each pass costs budget. Stop as soon as you find relevant results. Do not run all passes if pass 1 succeeds.