Tutorial — Authored Answers for Top Questions

Augmented Traveler local tutorial. Back to AI Mind — Tutorials · overview: Tutorial — How the Mind Works · animation: Tutorial — Sequencer for Monologues · in Unreal: Tutorial — Authoring a Q&A Experience with Ambient Idle

Most visitors ask the same questions. For those we write, record and animate an authored answer and play it instead of generating one. This page covers choosing the questions, making the answers, the matching rules, answer IDs, fallbacks, and updating from logs.

Draft. Written from AI Mind — Process, the Sentence Transformers docs and the Unreal tutorials. The threshold numbers are starting points to tune, not published values. Conversation designer + AI conversation engineer; historian approves every script.

1. Why authored first

• Best performance. Recorded by the actor, directed, animated with full body.
• Reviewed. The historian approves every word.
• Fast. No LLM, no TTS. Playback starts at once.
• Offline-safe. The same clips carry the offline mode (Tutorial — Authoring a Timed Experience).

2. Choosing the top questions

Before opening:

1. Take the golden set (Tutorial — Evaluating Answers).
2. Rank by how often staff and school groups expect each question.
3. Pick 30–60 for the first release. Include the "hello", "who are you" and "goodbye" beats.

After opening:

1. Weekly, group logged questions by the matched answer or by nearest neighbour.
2. Any generated-answer cluster above ~1% of questions is a candidate.
3. Any cluster with poor blind-review scores is a priority, even if small.

3. Writing the answer

• Length: 20–45 seconds spoken. One idea.
• Sources: every sentence cites a Corpus A or B ID in the script margin.
• Ending: a hook to a related question, e.g. "Ask me how the Danish king heard of it."
• Variants: write 2 short alternates for the most common answers so repeat visitors hear something new.
• Historian sign-off on the script before recording.

4. Recording and animating

1. Record with the contracted actor under the voice director (Voice — Process). Log the take and voice version.
2. Capture body and face performance, or keyframe gesture beats (Mocap — Process).
3. Assemble a Level Sequence per answer in Sequencer (Tutorial — Sequencer for Monologues). Name it LS_AA_MM_0012.
4. Captions from the approved script, timed to the take (Tutorial — Captions UI).
5. Check blend in and out of the idle loop (Tutorial — Authoring a Q&A Experience with Ambient Idle).

5. Answer IDs

AA-<figure>-<nnnn>[-v<variant>], e.g. AA-MM-0012-v2.

Each ID has a record:

id: AA-MM-0012
intent: comet_discovery
program: [comet-1847]
paraphrases:
  - How did you find the comet?
  - Tell me about discovering the comet
  - What did you see on the roof that night?
script_ids: [MM-A-1847-10-01-a]
sequence: LS_AA_MM_0012
audio_take: MM_VO_0012_t3
historian_status: approved
version: 3

The service sends this ID in answer_meta with authored: true (Tutorial — Conversation Client Plugin).

6. Matching rules

A question plays an authored answer only when all of these pass:

1. Intent match. The intent classifier's label equals the answer's intent.
2. Similarity. Cosine similarity between the question and the best paraphrase is ≥ 0.80 (start value; tune on the golden set).
3. Margin. Best score minus the second-best other answer is ≥ 0.05, so near-ties go to retrieval.
4. Program. The answer is valid for the current program and cutoff.
5. Policy. Input rails passed. A policy topic always goes to a canned reply, never an authored answer.

Model choice. This is symmetric matching (short question vs short paraphrase). Sentence Transformers' docs say to pick a model suited to the task; test the same embedding model as retrieval first.

Skeleton — test before use:

import numpy as np

SIM_MIN = 0.80
MARGIN_MIN = 0.05


def match(q_vec, intent, program, answers):
    """answers: list of dicts with id, intent, program, para_vecs (n x d, normalised)."""
    q = q_vec / np.linalg.norm(q_vec)
    scored = []
    for a in answers:
        if program not in a["program"]:
            continue
        scored.append((float(np.max(a["para_vecs"] @ q)), a))
    if not scored:
        return None
    scored.sort(key=lambda s: s[0], reverse=True)
    best_score, best = scored[0]
    second = scored[1][0] if len(scored) > 1 else 0.0
    if best["intent"] != intent:
        return None
    if best_score < SIM_MIN or best_score - second < MARGIN_MIN:
        return None
    return best["id"]

7. Fallbacks

8. Updating after logs

• Weeks 1–12 after opening: historian and conversation designer read all logs weekly (Ground Rules — Historical Virtual Humans).
• Add paraphrases from real questions that should have matched. Re-run the golden set.
• Remove paraphrases that caused wrong matches.
• New answers go through script → sign-off → recording → Sequencer → ID record.
• Urgent fixes: pin a canned reply overnight on the dashboard until the new answer is recorded (Tutorial — Staff Dashboard).
• Version every change to the ID record; the log stores the version played.

9. Practice exercise

1. Write the ID record for "What was the gold medal?" with eight paraphrases.
2. Write five near-miss questions (e.g. "Did you win any prizes?", "Who gave you money?").
3. Run the matcher. Adjust paraphrases until the near-misses route correctly.

10. Check before a venue build

• 30–60 authored answers approved, recorded and in Sequencer
• Every answer has an ID record with intent, paraphrases and sources
• Threshold and margin tuned on the golden set; false matches < 2%
• Fallbacks tested with the network unplugged
• Log shows answer ID and version for every authored play

Sources

• Sentence Transformers, Semantic search: https://sbert.net/examples/sentence_transformer/applications/semantic-search/README.html
• Internal: AI Mind — Process, Tutorial — Sequencer for Monologues, Tutorial — Conversation Client Plugin
• Full list: Sources — 2026-09-20