Translation Pipeline

Every Falara translation job passes through a multi-agent pipeline of specialized AI agents. Each agent has a distinct responsibility, and agents communicate via structured handoffs.

---

Pipeline Flow

Input Text / File
      │
      ▼
┌─────────────┐
│  Supervisor  │  Validates briefing, segments text, plans context
└──────┬──────┘
       │
       ▼  (optional)
┌──────────────┐
│ Source Review │  Checks source text quality (spelling, grammar, consistency)
└──────┬───────┘
       │
       ▼
┌─────────────┐
│  Translator  │  Translates segments sequentially with rolling context
└──────┬──────┘
       │
       ▼
┌──────┐
│  QA  │  Scores each segment (0–100) against source
└──┬───┘
   │
   ├── Score ≥ 95 ──→ ✅ completed
   │
   └── Score < 95 ──→ Correction loop (max 2×)
                             │
                             ├── Score ≥ 80 after loops ──→ ✅ completed
                             │
                             └── Score < 80 after loops ──→ ⚠️ needs_review

---

Agents

1. Supervisor

• Validates the briefing (language codes, mode, glossary, constraints)
• Segments the input text into translation units
• Builds a context plan (prior segments passed to Translator for coherence)
• Detects and blocks untranslatable content

Skipped in batch mode: When texts is provided (list of short strings), each string becomes a segment directly without Supervisor segmentation.

2. Source Review (optional)

• Checks the source text for spelling, grammar, punctuation, and terminology consistency
• Outputs source_error delivery notes — purely informational, does not block translation
• Does not modify the job status
• Activated by source_review: true in the request, or by tier default (Business: on by default, Enterprise: always on)

3. Translator

• Translates each segment using the source context and prior translated segments
• Applies glossary terms and respects constraints
• Uses prompt caching for efficiency on large documents
• Runs with rolling context: each segment receives the previous segment's translation as context

4. QA

• Independently evaluates each translated segment on a 0–100 scale
• Criteria: source fidelity, language quality, terminology adherence, constraint compliance
• Segments below the pass threshold trigger a correction loop

5. Correction Loops

• Maximum 2 correction loops per job
• Only segments below threshold are re-translated in each loop
• After correction, QA re-evaluates the corrected segments
• If score ≥ 80 after all loops: job → completed
• If score < 80 after all loops: job → needs_review with delivery notes

---

Thresholds

Threshold	Value	Description
QA pass score	95	Segment passes QA without correction
Min acceptable score	80	Minimum to reach completed status
Global segment threshold	90	Default minimum score
Short segment (≤7 words)	85	Relaxed threshold
Very short segment (≤3 words)	75	Highly relaxed threshold

---

Job Status Progression

queued → processing → [correcting →] completed
                                   → completed_with_blocks
                                   → needs_review
                    → failed
                    → dead