> For a complete page index, fetch https://docs.synthflow.ai/llms.txt. For full documentation content, fetch https://docs.synthflow.ai/llms-full.txt.

# Composite Multilingual NLP in Synthflow

> Synthflow's composite NLP architecture combines rule-based techniques, machine learning models, and large language models to deliver accurate multilingual conversational AI across 40+ languages.

Synthflow employs a **composite multilingual NLP architecture** that combines multiple natural language processing techniques to deliver accurate, context-aware conversational AI. Rather than relying on a single approach, Synthflow layers rule-based systems, machine learning models, and large language models (LLMs) to create robust, production-ready agents.

## What is Composite NLP?

**Composite NLP** refers to the integration of multiple complementary techniques in a processing pipeline:

* **Rule-based systems**: Explicit logic for pronunciation, content filtering, and deterministic behavior
* **Machine learning models**: ASR (Automatic Speech Recognition) and TTS (Text-to-Speech) for audio processing
* **Large language models**: GPT-4.1, GPT-5.1, GPT-5.2 and Synthflow-optimized models for natural language understanding and generation
* **Retrieval systems**: Semantic search and RAG (Retrieval-Augmented Generation) for knowledge access

This multi-layer composition ensures that each component handles what it does best, resulting in higher accuracy, better control, and more reliable performance than any single technique alone.

***

## Multilingual Support

Synthflow supports **40+ languages** with seamless multilingual mode, enabling agents to understand and respond in the user's preferred language automatically.

### Supported Languages

Synthflow agents can converse fluently in over 40 languages, including:

* **European**: English, Spanish, French, German, Italian, Portuguese, Dutch, Polish, Russian, Ukrainian, Swedish, Norwegian, Danish, Finnish, Czech, Romanian, Greek, Turkish
* **Asian**: Chinese (Mandarin), Japanese, Korean, Hindi, Bengali, Tamil, Telugu, Thai, Vietnamese, Indonesian, Malay, Tagalog
* **Middle Eastern**: Arabic, Hebrew, Persian (Farsi)
* **Other**: Swahili, Zulu, Afrikaans

### Multilingual Mode

Agents can automatically detect and respond in the user's language without manual configuration:

* **Automatic language detection**: Identifies the user's language from their first utterance
* **Seamless switching**: Maintains context when users switch languages mid-conversation
* **Consistent personality**: Agent personality and behavior remain consistent across languages
* **Cultural adaptation**: Responses adapt to cultural norms and communication styles

### How Multilingual Processing Works

Synthflow's composite NLP pipeline processes multilingual conversations through multiple specialized components:

1. **ASR (Deepgram)**: Converts speech to text in the detected language
2. **LLM Processing**: Understands intent and generates responses in the appropriate language
3. **TTS (ElevenLabs)**: Converts text responses back to natural-sounding speech
4. **Rule-based Overlays**: Applies custom vocabulary and filter words regardless of language

This architecture ensures consistent quality across all supported languages.

### Documentation

* [About Supported Languages](/about-supported-languages) - Complete language list and capabilities

***

## Composite NLP Architecture

Synthflow's NLP system combines multiple techniques in a sophisticated processing pipeline. Each layer contributes specific capabilities that work together to create natural, accurate conversations.

### Layer 1: Automatic Speech Recognition (ASR)

**Technology**: Synthflow STT and choice of other providers

**Purpose**: Convert spoken audio to text with high accuracy across languages

**Capabilities**:

* Real-time transcription with low latency
* Multilingual support for 40+ languages
* Accent and dialect recognition
* Background noise filtering
* Punctuation and formatting

The ASR layer transforms user speech into text that subsequent layers can process.

### Layer 2: Large Language Models (LLMs)

**Technology**: OpenAI GPT models, Gemini models, Synthflow-optimized models

**Purpose**: Understand user intent, generate contextually appropriate responses, and manage conversation flow

**Available Models**:

| Model            | Best For                    | Characteristics                                                                         |
| ---------------- | --------------------------- | --------------------------------------------------------------------------------------- |
| **GPT-4.1**      | Complex tasks               | Well-suited when latency is less critical                                               |
| **GPT-4.1 Nano** | Speed and cost              | Lightweight variant of GPT-4.1                                                          |
| **GPT-4.1 Mini** | Cost optimization           | Similar reasoning with lower cost                                                       |
| **GPT-5 Chat**   | Cutting-edge quality        | Advanced conversational capabilities                                                    |
| **GPT-5 Mini**   | Balanced performance        | Compact GPT-5 variant balancing performance and cost                                    |
| **GPT-5 Nano**   | Speed and cost              | Lightweight GPT-5 variant                                                               |
| **GPT-5.1**      | Low latency                 | Advanced model optimized for speed                                                      |
| **GPT-5.2**      | Concise, structured prompts | Defaults to brevity; prefers sectioned prompts and explicit persistence/efficiency cues |
| **Synthflow**    | Voice-specific use cases    | Optimized for phone conversations                                                       |

Users can select the appropriate model based on their latency, cost, and quality requirements.

**LLM Capabilities**:

* Natural language understanding across languages
* Context retention throughout conversations
* Intent recognition and entity extraction
* Response generation with personality consistency
* Complex reasoning and decision-making

### Layer 3: Text-to-Speech (TTS)

**Technology**: ElevenLabs and other providers

**Purpose**: Convert generated text responses into natural-sounding speech

**Capabilities**:

* Natural prosody and intonation
* Emotional expression
* Multiple voice options
* Multilingual voice synthesis
* Real-time generation with low latency

The TTS layer ensures agents sound natural and engaging across all supported languages.

### Layer 4: Rule-Based Enhancements and Hybrid Logic

While LLMs handle most processing, Synthflow overlays rule-based systems for precise control and offers hybrid approaches that combine both techniques:

#### Flow Designer: Deterministic and Natural Language Transition Conditions

**Purpose**: Flexible conversation routing combining rule-based and ML-based logic

**How it works**: The Flow Designer supports both deterministic (rule-based) and natural language (ML-based) transition conditions, allowing users to choose the right approach for each decision point:

**Deterministic (Rule-Based) Conditions**:

* Exact variable matching: `if {user_response} == "yes"`
* Numeric comparisons: `if {age} >= 18`
* Boolean logic: `if {is_member} == true`
* List membership: `if {selected_option} in ["A", "B", "C"]`
* Empty/null checks: `if {email} is not empty`

**Natural Language (ML-Based) Conditions**:

* Intent detection: "If the user wants to speak to a human"
* Semantic understanding: "If the user expresses frustration or anger"
* Flexible matching: "If the user asks about pricing, costs, or fees"
* Context-aware: "If the user mentions a medical emergency"

**Hybrid Approach Benefits**:

* **Precision where needed**: Use deterministic rules for compliance-critical decisions
* **Flexibility where helpful**: Use natural language for handling varied user expressions
* **Best of both worlds**: Combine rule-based reliability with ML-based adaptability

**Example: Appointment Booking Flow**

```
Deterministic condition: if {appointment_date} is in the past → Error message
Natural language condition: if user wants to reschedule → Jump to reschedule flow
Deterministic condition: if {time_slot} == "unavailable" → Offer alternatives
Natural language condition: if user expresses urgency → Prioritize next available slot
```

This composite approach allows conversation designers to leverage rule-based logic for structured decisions while using ML-based natural language understanding for flexible, human-like interactions.

**Documentation**:

* [Configure Branch Nodes](/configure-branch-nodes) - Deterministic and natural language conditions
* [Flow Designer](/flow-designer) - Hybrid logic patterns

#### Custom Vocabulary

**Purpose**: Ensure correct pronunciation of business-specific terms

**How it works**: Users define custom pronunciations for:

* Brand names (e.g., "Synthflow" pronounced correctly)
* Product codes or SKUs
* Technical terminology
* Industry-specific jargon
* Names and proper nouns

The system applies these rules before TTS processing, ensuring consistent pronunciation regardless of the underlying model.

**Configuration**: Low-code/no-code interface in General Configuration settings

**Documentation**:

* [General Configuration](/general-configuration) - Custom Vocabulary setup

#### Filter Words

**Purpose**: Content filtering and guardrails

**How it works**: Rule-based blocking of specific terms the agent should never speak:

* Sensitive information placeholders
* Inappropriate language
* Competitor names
* Confidential terms

This rule-based layer acts as a safety net, preventing the LLM from generating unwanted content.

**Configuration**: Simple list-based interface in General Configuration

**Documentation**:

* [General Configuration](/general-configuration) - Filter Words configuration

### Layer 5: Retrieval-Augmented Generation (RAG)

**Purpose**: Combine retrieval (rule-based search) with generation (ML/LLM) for accurate, grounded responses

**How RAG Works**:

1. **Document Storage**: Upload PDFs, web pages, or text documents to Knowledge Base
2. **Semantic Search**: When users ask questions, the system searches for relevant content using embeddings
3. **Context Injection**: Retrieved information is injected into the LLM prompt
4. **Grounded Generation**: LLM generates responses based on retrieved facts, not just training data

**Benefits**:

* **Accuracy**: Responses grounded in your specific documents and data
* **Up-to-date information**: Reference current content without retraining models
* **Source attribution**: Responses based on verifiable sources
* **Reduced hallucinations**: LLM constrained by retrieved facts

**RAG Architecture Components**:

* **Embeddings**: Vector representations of document chunks for semantic search
* **Retrieval**: Rule-based and semantic search to find relevant content
* **Ranking**: ML-based relevance scoring
* **Generation**: LLM synthesis of retrieved information into natural responses

This composite approach (retrieval + generation) outperforms either technique alone.

### Documentation

* [Knowledge Base](/knowledge-base) - RAG architecture explanation
* [General Configuration](general-configuration) - Model selection, Custom Vocabulary, Filter Words

***

## Composite Techniques in Practice

### Example: Customer Support Agent

A customer support agent demonstrates how multiple NLP techniques work together:

**User**: "What's your return policy for electronics?"

**Processing Pipeline**:

1. **ASR (Deepgram)**: Converts speech to text: "What's your return policy for electronics?"

2. **Custom Vocabulary (Rule-based)**: Recognizes "electronics" as a product category (no pronunciation adjustment needed)

3. **RAG Retrieval (Semantic Search)**: Searches Knowledge Base for relevant policy documents
   * Finds: "Electronics Return Policy.pdf"
   * Extracts relevant sections about 30-day returns, condition requirements

4. **LLM (GPT-4.1)**: Generates response based on retrieved policy:
   * Understands question intent
   * Synthesizes policy information into conversational response
   * Maintains agent personality and tone

5. **Filter Words (Rule-based)**: Checks response against blocked terms (none found)

6. **TTS (ElevenLabs)**: Converts response to natural speech

**Agent Response**: "We offer a 30-day return policy for electronics. Items must be in original packaging and in like-new condition. Would you like me to email you the complete policy details?"

This example shows how rule-based systems (Custom Vocabulary, Filter Words), machine learning (ASR, TTS), semantic search (RAG retrieval), and LLMs work together in a composite pipeline.

***

## Advanced Composite NLP Features

### Sentiment Detection

**Technology**: Proprietary orchestration with ML models

**Purpose**: Detect user emotions and adjust agent behavior

**How it works**:

* Analyzes user speech patterns, word choice, and tone
* Identifies frustration, satisfaction, urgency, confusion
* Triggers escalation rules when negative sentiment detected
* Adjusts agent responses to be more empathetic

**Use case**: Automatically transfer frustrated customers to human agents

### Fallback Logic

**Technology**: Rule-based with ML confidence scoring

**Purpose**: Handle situations when the LLM is uncertain

**How it works**:

* LLM provides confidence scores for its responses
* Rule-based thresholds trigger fallback behaviors
* Options: ask clarifying questions, transfer to human, use default responses

**Use case**: Prevent agents from guessing when they don't understand

### Escalation Triggers

**Technology**: Composite rule-based and ML detection

**Purpose**: Identify when human intervention is needed

**Triggers**:

* Sentiment analysis detects frustration (ML)
* User explicitly requests human agent (rule-based)
* Conversation exceeds time threshold (rule-based)
* LLM confidence below threshold (ML scoring)
* Specific keywords detected (rule-based)

**Use case**: Seamless handoff to human agents when AI reaches its limits

### Documentation

* [AI Transparency Statement](/ai-transparency) - Complete composite AI architecture details

***

## Model Selection and Optimization

Synthflow's composite approach extends to **model selection**, allowing users to choose the right LLM for their specific requirements.

### Choosing the Right Model

**For low latency requirements**:

* GPT-5.1: Advanced model optimized for speed
* GPT-4.1 Nano / GPT-5 Nano: Lightweight variants for fastest responses
* Synthflow: Purpose-built for voice conversations

**For complex reasoning**:

* GPT-4.1: Deep understanding and complex task handling
* GPT-5 Chat: Cutting-edge conversational quality
* GPT-5.2: Most advanced model for structured prompts

**For cost optimization**:

* GPT-4.1 Mini / GPT-5 Mini: Similar capabilities at lower cost
* GPT-4.1 Nano / GPT-5 Nano: Lightest-weight options

**For general use (recommended)**:

* GPT-4.1: Strong all-round performance for most use cases

### Specialized Synthflow Models

Synthflow develops **specialized models** for voice-specific use cases:

**Voicemail Detection Model**:

* Purpose-built to identify when calls reach voicemail
* Higher accuracy than general-purpose LLMs
* Enables appropriate voicemail message delivery
* Prevents wasted conversation attempts

**Customer Data Input Model**:

* Optimized for extracting structured data from conversations
* Better accuracy for phone numbers, dates, addresses
* Reduces errors in data collection

These specialized models demonstrate Synthflow's composite approach: using the right model for each specific task rather than one-size-fits-all.

### Documentation

* [General Configuration](/general-configuration) - AI model selection
* [Model-specific prompting notes](/prompting-best-practices#model-specific-notes) - Model-specific optimization

***

## Customizable NLP Pipelines

Synthflow's composite architecture is **customizable**, allowing users to configure the NLP pipeline for their specific needs.

### Configuration Options

**Model Selection**: Choose which LLM powers your agent's responses

**Custom Vocabulary**: Add rule-based pronunciation overrides

**Filter Words**: Define rule-based content blocking

**Knowledge Base**: Configure RAG retrieval parameters:

* Trigger conditions for knowledge base searches
* Number of relevant chunks to retrieve
* Confidence thresholds for using retrieved information

**Voice Selection**: Choose TTS voice and style

**Sentiment Thresholds**: Configure when sentiment triggers escalation

### Predefined and Custom Pipelines

**Predefined Pipelines**: Synthflow provides optimized default configurations for common use cases:

* Customer support
* Sales and lead qualification
* Appointment scheduling
* Information lookup

**Custom Pipelines**: Advanced users can customize the processing pipeline:

* Adjust RAG retrieval parameters
* Configure custom fallback logic
* Define escalation rules
* Set model-specific parameters

This flexibility ensures the composite NLP system adapts to diverse business requirements.

***

## Benefits of Composite Multilingual NLP

### 1. Higher Accuracy

By combining techniques, Synthflow achieves higher accuracy than any single approach:

* Rule-based systems handle cases requiring precision
* ML models handle complex pattern recognition
* LLMs provide natural language understanding
* RAG grounds responses in factual data

### 2. Better Control

Composite architecture provides control at multiple levels:

* Rule-based layers for deterministic behavior
* Model selection for performance tuning
* RAG for content control
* Fallback logic for edge cases

### 3. Multilingual Consistency

The same composite pipeline works across 40+ languages:

* ASR handles multilingual speech recognition
* LLMs understand and generate in multiple languages
* Rule-based enhancements (Custom Vocabulary, Filter Words) apply universally
* TTS produces natural speech in any supported language

### 4. Reduced Hallucinations

Multiple techniques work together to prevent incorrect responses:

* RAG grounds responses in retrieved facts
* Rule-based filters block unwanted content
* Confidence scoring triggers fallbacks when uncertain
* Sentiment detection catches when conversations go off-track

### 5. Flexibility and Scalability

Composite architecture allows:

* Swapping models as technology improves
* Adding new techniques without rebuilding
* Scaling different components independently
* Optimizing cost vs. performance trade-offs

***

## Summary

Synthflow's **composite multilingual NLP** represents a sophisticated approach to conversational AI:

**Key Components**:

* **ASR (Deepgram)**: Speech-to-text across 40+ languages
* **LLMs (OpenAI, Synthflow)**: Natural language understanding and generation
* **TTS (ElevenLabs)**: Natural speech synthesis
* **Rule-based Systems**: Custom Vocabulary, Filter Words, fallback logic
* **RAG**: Retrieval-augmented generation for grounded responses
* **Proprietary Orchestration**: Sentiment detection, escalation triggers, confidence scoring

**Composite Approach Benefits**:

* ✅ Higher accuracy through complementary techniques
* ✅ Better control with rule-based overlays
* ✅ Multilingual support across 40+ languages
* ✅ Reduced hallucinations via RAG and confidence scoring
* ✅ Flexible, customizable pipelines
* ✅ Specialized models for specific use cases

By combining rule-based techniques, machine learning models, and large language models in a sophisticated processing pipeline, Synthflow delivers enterprise-grade conversational AI that is accurate, controllable, and reliable across languages and use cases.

### Documentation

* [About Supported Languages](/about-supported-languages) - Multilingual capabilities
* [General Configuration](/general-configuration) - Model selection and composite features
* [Knowledge Base](/knowledge-base) - RAG architecture
* [AI Transparency Statement](/ai-transparency) - Complete composite AI system details