15.1 Objective
Until now, ACAI has primarily been described around text-based intelligence.
A real multimodal AI system must be able to work with:
Text
Images
Audio
Video
Documents
and combine information from multiple modalities.
The goal is:
TEXT
IMAGE
AUDIO
VIDEO
DOCUMENT
│
▼
MULTIMODAL UNDERSTANDING
│
▼
REASONING
│
▼
ACTION
│
▼
VERIFICATION
│
▼
RESULT
This transforms ACAI from a primarily text-oriented architecture into a multimodal AI platform.
15.2 Multimodal Architecture
The high-level design becomes:
USER
│
┌────────────────┼────────────────┐
▼ ▼ ▼
TEXT IMAGE AUDIO
│ │ │
└────────────────┼────────────────┘
│
▼
VIDEO
│
▼
DOCUMENTS
│
▼
INPUT PROCESSING
│
▼
MULTIMODAL ENCODING
│
▼
MULTIMODAL MODEL
│
┌─────────┼─────────┐
▼ ▼ ▼
REASONING RETRIEVAL TOOLS
│ │ │
└─────────┼─────────┘
▼
VERIFIER
│
▼
OUTPUT
15.3 Input Layer
Every modality should enter through a controlled input layer.
For example:
POST /input/text
POST /input/image
POST /input/audio
POST /input/video
POST /input/document
The system then normalizes these inputs into internal representations.
15.4 Input Normalization
Different media have different formats.
Examples:
Image:
PNG
JPEG
WEBP
Audio:
WAV
MP3
FLAC
Video:
MP4
MOV
WEBM
Documents:
PDF
DOCX
TXT
HTML
The input layer should validate and normalize supported formats before processing.
15.5 Image Understanding
A vision pipeline can look like:
IMAGE
↓
Validation
↓
Preprocessing
↓
Vision Encoder / Vision Model
↓
Visual Representation
↓
Reasoning
Possible tasks include:
Image description
Object understanding
Document image analysis
Visual question answering
Scene understanding
Image comparison
Chart interpretation
15.6 Image Preprocessing
Before sending an image to a model:
Upload
↓
Format validation
↓
Resolution check
↓
Resize if required
↓
Metadata handling
↓
Model input
The exact preprocessing depends on the selected model.
15.7 OCR
OCR means Optical Character Recognition.
A document-image workflow:
IMAGE
↓
OCR
↓
TEXT
↓
STRUCTURE
↓
RETRIEVAL
↓
REASONING
For example, a scanned document may contain text that is not directly machine-readable.
OCR converts visual characters into machine-readable text.
15.8 Document Intelligence
Documents should not always be treated as plain text.
A document may contain:
Title
Paragraphs
Tables
Images
Headers
Footnotes
Page numbers
Forms
Signatures
Charts
A document-processing pipeline can preserve these structures.
DOCUMENT
↓
PARSER
↓
LAYOUT ANALYSIS
↓
TEXT + TABLES + IMAGES
↓
CHUNKING
↓
EMBEDDING
↓
INDEX
15.9 PDF Processing
A PDF may contain either:
Machine-readable text
or:
Scanned images
Therefore:
PDF
│
├── Text available
│ ↓
│ Extract
│
└── No text
↓
OCR
The system should choose the appropriate path.
15.10 Document Chunking
Large documents should generally be divided into manageable retrieval units.
Example:
Document
↓
Chapter
↓
Section
↓
Paragraph
↓
Chunk
Each chunk can store metadata:
{
"document_id": "doc_001",
"page": 42,
"section": "Architecture",
"text": "..."
}
This improves traceability.
15.11 Multimodal Embeddings
Traditional embeddings may represent text.
A multimodal system can represent:
Text
Image
Audio
Video
in compatible or related embedding spaces, depending on the model architecture.
Conceptually:
TEXT ───► EMBEDDING
IMAGE ──► EMBEDDING
AUDIO ──► EMBEDDING
VIDEO ──► EMBEDDING
This allows cross-modal retrieval in systems designed for it.
15.12 Cross-Modal Retrieval
Example:
User:
"Find the image that contains the same diagram discussed in this document."
Possible pipeline:
Text Query
↓
Text Embedding
↓
Multimodal Search
↓
Image Candidates
↓
Ranking
↓
Result
This is more powerful than text-only search.
15.13 Image-to-Text Retrieval
Example:
IMAGE
↓
Visual Representation
↓
Search
↓
Relevant Documents
A user could upload an image and ask:
"What documents are related to this image?"
The system can search indexed multimodal representations.
15.14 Text-to-Image Retrieval
Reverse direction:
TEXT
↓
Embedding
↓
Image Search
↓
Relevant Images
This can support:
Media libraries
Research archives
Design systems
Product catalogs
Document repositories
15.15 Audio Intelligence
Audio processing can contain several stages:
AUDIO
↓
Preprocessing
↓
Speech Detection
↓
Speech Recognition
↓
Text
↓
Reasoning
For speech:
Speech
↓
ASR
↓
Transcript
ASR means Automatic Speech Recognition.
15.16 Audio Beyond Speech
Audio can contain more than spoken language.
Potential analysis includes:
Speech
Music
Environmental sounds
Speaker characteristics
Silence
Audio events
The supported capabilities depend on the selected models.
15.17 Speech-to-Text
Example:
USER SPEAKS
↓
MICROPHONE
↓
AUDIO STREAM
↓
ASR
↓
TRANSCRIPT
↓
LLM
↓
ANSWER
This enables voice interaction with ACAI.
15.18 Text-to-Speech
The reverse pipeline:
USER
↓
TEXT
↓
LLM
↓
RESPONSE TEXT
↓
TTS
↓
AUDIO
↓
USER
TTS means Text-to-Speech.
15.19 Voice Assistant
A complete voice interaction can therefore be:
USER SPEECH
↓
ASR
↓
UNDERSTANDING
↓
MEMORY
↓
PLANNING
↓
TOOLS
↓
REASONING
↓
TEXT RESPONSE
↓
TTS
↓
VOICE
The same orchestration architecture can be reused.
15.20 Video Understanding
Video introduces a time dimension.
A simple pipeline:
VIDEO
↓
Metadata
↓
Frame Sampling
↓
Visual Analysis
↓
Audio Extraction
↓
Speech Recognition
↓
Temporal Reasoning
↓
Summary
15.21 Video Is Not Just Many Images
Suppose a video contains:
Frame 1:
Person enters.
Frame 2:
Person picks up object.
Frame 3:
Person leaves.
Understanding the sequence requires temporal context.
Therefore:
Frame understanding
+
Temporal reasoning
is necessary for many video tasks.
15.22 Video Sampling
Processing every frame may be expensive.
Instead:
VIDEO
↓
Sampling Strategy
↓
Selected Frames
↓
Vision Model
Sampling frequency depends on the task.
For example, a static presentation may require fewer frames than a fast-action video.
15.23 Keyframe Extraction
The system can identify important frames:
Video
↓
Scene changes
↓
Keyframes
↓
Analysis
Keyframes reduce redundant processing.
15.24 Video + Audio
A better video understanding pipeline combines both:
VIDEO
│
┌────────┴────────┐
▼ ▼
FRAMES AUDIO
│ │
▼ ▼
VISION ASR
│ │
└────────┬────────┘
▼
MULTIMODAL
REASONING
This allows ACAI to combine what was seen with what was said.
15.25 Example Multimodal Request
User:
"Summarize this lecture video and identify the three diagrams shown."
ACAI could execute:
Video
↓
Audio extraction
↓
Transcript
↓
Frame sampling
↓
Diagram detection
↓
Visual understanding
↓
Transcript + visuals
↓
Reasoning
↓
Final summary
15.26 Multimodal Memory
ACAI memory can store more than text.
Possible memory records:
Text memory
Image memory
Audio transcript
Video event
Document reference
For example:
{
"memory_id": "mem_001",
"type": "image",
"description": "...",
"embedding": "...",
"source": "user_upload"
}
15.27 Media Metadata
Every media object should have metadata.
Example:
{
"media_id": "img_001",
"type": "image",
"mime_type": "image/jpeg",
"size": 2048000,
"created_at": "...",
"owner_id": "user_001"
}
For video:
{
"media_id": "vid_001",
"type": "video",
"duration": 120,
"width": 1920,
"height": 1080
}
15.28 Object Storage
Large files should generally not be stored directly inside relational database rows.
Use:
Application
↓
Object Storage
↓
Media File
Database stores:
metadata
permissions
location/reference
processing status
15.29 Media Processing Queue
Media processing can be expensive.
Therefore:
UPLOAD
↓
JOB CREATED
↓
QUEUE
↓
MEDIA WORKER
↓
PROCESSING
↓
STORAGE
↓
INDEX
This matches the production architecture from Chapter 13.
15.30 Processing States
A media object can have:
uploaded
validating
processing
indexed
ready
failed
Example:
uploaded
↓
processing
↓
ready
If something fails:
processing
↓
failed
15.31 Multimodal Agent
The agent can now choose among:
Text tool
Vision tool
OCR tool
Audio tool
Video tool
Search
Database
Calculator
Example:
User
↓
"Look at this chart and calculate the percentage change."
↓
Vision
↓
Extract values
↓
Calculator
↓
Verifier
↓
Answer
This demonstrates why multimodal systems benefit from tool orchestration.
15.32 Vision + Tool Calling
A model should not necessarily perform exact arithmetic itself when a calculator is available.
Pipeline:
IMAGE
↓
VISION
↓
VALUES
↓
CALCULATOR
↓
RESULT
↓
VERIFIER
This reduces avoidable numerical errors.
15.33 Document + Retrieval + Reasoning
Example:
PDF
↓
Parser
↓
OCR if needed
↓
Chunks
↓
Embeddings
↓
Vector Index
↓
User Question
↓
Retrieval
↓
Relevant Sections
↓
Reasoning
↓
Citation / Source Reference
The final answer should distinguish between information found in the document and model-generated interpretation.
15.34 Multimodal Verification
Verification can happen at multiple stages.
Input verification
↓
Processing verification
↓
Model-output verification
↓
Tool verification
↓
Final verification
For example:
Image says:
$500
OCR says:
$800
The system should flag a possible extraction conflict instead of confidently returning an unverified value.
15.35 Confidence
Multimodal systems may produce confidence signals.
For example:
OCR confidence
Vision confidence
Speech recognition confidence
Retrieval score
These should be treated as signals, not universal guarantees of correctness.
15.36 Uncertainty Handling
If ACAI cannot reliably interpret an image:
Confidence low
↓
Do not pretend certainty
↓
Ask for clearer image
Similarly:
Audio unclear
↓
Request repetition
Good multimodal behavior includes knowing when the input is insufficient.
15.37 Privacy
Multimedia can contain highly sensitive information.
Examples:
Faces
Voices
Documents
Locations
Personal conversations
Private recordings
Therefore the system needs:
Access controls
Encryption
Retention policies
Deletion mechanisms
Audit logs
Data collection and model training should also follow applicable consent and privacy requirements.
15.38 Media Access Control
A media object should be associated with an owner or authorized scope.
User A
↓
Media A
must not automatically be accessible to:
User B
Authorization should be checked before:
Download
View
Process
Share
Delete
15.39 Multimodal API
A conceptual request:
{
"input": [
{
"type": "text",
"content": "Analyze this image."
},
{
"type": "image",
"media_id": "img_001"
}
]
}
The orchestrator determines the appropriate processing path.
15.40 Unified Input Representation
Internally, ACAI can normalize different inputs into a common structure:
class MultimodalInput:
type: str
content: str | bytes | dict
metadata: dict
Possible types:
text
image
audio
video
document
This allows the orchestration layer to operate consistently.
15.41 Multimodal Workflow Example
Request:
"Read this PDF, inspect the chart, and explain the trend."
Workflow:
USER
↓
PDF
↓
DOCUMENT PROCESSOR
↓
TEXT + IMAGE
↓
CHART DETECTION
↓
VISION
↓
RETRIEVAL
↓
REASONING
↓
VERIFIER
↓
ANSWER
15.42 Multimodal Search Architecture
SEARCH QUERY
│
┌──────────┼──────────┐
▼ ▼ ▼
TEXT IMAGE AUDIO
│ │ │
└──────────┼──────────┘
▼
MULTIMODAL INDEX
│
▼
RANKING
│
▼
RESULTS
15.43 Media Index
A media index might contain:
media_id
type
embedding
metadata
owner_id
document_id
timestamp
source
For video, additional information may include:
timestamp
scene
frame
transcript segment
15.44 Temporal Indexing
Video information can be indexed by time:
00:00–00:30
00:30–01:00
01:00–01:30
Then ACAI can answer:
"What happened around 02:15?"
by retrieving the relevant temporal segment.
15.45 Multimodal RAG
Traditional RAG:
Question
↓
Text Retrieval
↓
Context
↓
LLM
Multimodal RAG:
Question
↓
Text + Image + Audio + Video Retrieval
↓
Multimodal Context
↓
Multimodal Model
↓
Answer
This enables richer research and document analysis.
15.46 Multimodal Research Assistant
A research workflow could be:
Research Question
↓
Web / Document Retrieval
↓
PDFs
↓
Charts
↓
Images
↓
Tables
↓
Audio / Video if relevant
↓
Cross-modal Analysis
↓
Evidence Organization
↓
Reasoning
↓
Final Report
The system should preserve source provenance so that conclusions can be traced back to supporting material.
15.47 Multimodal Creative System
ACAI can also support creative workflows:
Text Prompt
↓
Planning
↓
Image Generation
↓
Image Analysis
↓
Video Generation
↓
Audio / Voice
↓
Editing
↓
Quality Check
This connects multimodal intelligence with the creative-media capabilities envisioned for the broader platform.
15.48 End-to-End Multimodal Architecture
USER
│
┌───────────────────────┼────────────────────────┐
▼ ▼ ▼
TEXT IMAGE AUDIO
│ │ │
│ OCR/VISION ASR
│ │ │
└───────────────────────┼────────────────────────┘
▼
VIDEO
│
FRAME + AUDIO
│
▼
DOCUMENTS
│
PARSER / OCR
│
▼
MULTIMODAL PROCESSOR
│
▼
MEMORY / INDEX
│
▼
RETRIEVAL SYSTEM
│
▼
ORCHESTRATOR
│
┌───────────────┼────────────────┐
▼ ▼ ▼
PLANNER MODEL TOOLS
│ │ │
└───────────────┼────────────────┘
▼
VERIFIER
│
▼
OUTPUT
│
┌───────────┼───────────┐
▼ ▼ ▼
TEXT IMAGE AUDIO
│
▼
VIDEO
15.49 Implementation Order
Do not attempt to implement every modality simultaneously.
A practical order is:
STEP 1
Text foundation
STEP 2
Document processing
STEP 3
OCR
STEP 4
Image understanding
STEP 5
Audio transcription
STEP 6
Voice output
STEP 7
Video processing
STEP 8
Multimodal retrieval
STEP 9
Cross-modal reasoning
STEP 10
Multimodal agents
This reduces implementation complexity.
15.50 Testing Strategy
Each modality requires its own tests.
Image
Clear image
Low-resolution image
Multiple objects
Charts
Documents
Audio
Clear speech
Noise
Multiple speakers
Different accents
Silence
Video
Static scene
Fast scene
Long video
Multiple speakers
Audio + visual conflict
Documents
Text PDF
Scanned PDF
Tables
Images
Large document
Malformed document
15.51 Multimodal Regression Tests
Maintain a benchmark:
image_test_001
image_test_002
audio_test_001
video_test_001
document_test_001
Every major model or pipeline update should run against these tests.
15.52 Performance Optimization
Multimodal processing can be expensive.
Optimization techniques include:
Caching
Frame sampling
Image resizing
Batch processing
Asynchronous workers
Model routing
Result reuse
Selective modality processing
The principle is:
Process only what is necessary.
15.53 Cost-Aware Multimodal Routing
For a simple image:
Small vision model
For a difficult research task:
More capable multimodal model
For speech transcription:
Speech model
For exact arithmetic:
Calculator
Therefore the router should select the cheapest sufficiently capable component.
15.54 Failure Handling
Example:
Video uploaded
↓
Audio extraction fails
The system can still attempt:
Video frames
↓
Visual analysis
and clearly report that audio could not be processed.
This is another example of graceful degradation.
15.55 Chapter 15 Success Criteria
[✓] Multimodal architecture defined
[✓] Image processing defined
[✓] OCR defined
[✓] Document intelligence defined
[✓] Audio processing defined
[✓] Speech-to-text defined
[✓] Text-to-speech defined
[✓] Video processing defined
[✓] Frame sampling defined
[✓] Keyframe extraction defined
[✓] Multimodal embeddings defined
[✓] Cross-modal retrieval defined
[✓] Multimodal memory defined
[✓] Media storage defined
[✓] Media queues defined
[✓] Multimodal agents defined
[✓] Multimodal verification defined
[✓] Privacy architecture defined
[✓] Multimodal testing defined
[✓] Multimodal RAG defined
[✓] Multimodal research workflow defined
[✓] Multimodal creative workflow defined
15.56 Final Multimodal Vision
The complete concept is now:
ACAI
│
┌──────────────────┼──────────────────┐
▼ ▼ ▼
TEXT IMAGE AUDIO
│ │ │
└──────────────────┼──────────────────┘
▼
VIDEO
│
▼
DOCUMENTS
│
▼
MULTIMODAL UNDERSTANDING
│
▼
MEMORY
│
▼
RETRIEVAL
│
▼
REASONING
│
▼
ACTION
│
▼
VERIFICATION
│
▼
OUTPUT
ACAI is therefore no longer limited to understanding words. Its architecture can process different forms of information and connect them inside one reasoning workflow.
15.57 Next Chapter
Chapter 16 — Autonomous Agents, Long-Horizon Planning, Multi-Agent Systems, and Safe Execution
The next chapter will cover:
Agent architecture
Task decomposition
Long-horizon planning
Planning loops
Multi-agent systems
Specialized agents
Agent communication
Shared memory
Agent supervision
Human approval
Execution boundaries
Sandboxing
Tool permissions
Agent failure recovery
Goal verification
Long-running autonomous workflows
The target architecture becomes:
USER GOAL
↓
PLANNER
↓
TASK DECOMPOSITION
↓
┌───────────────┐
│ AGENT SYSTEM │
├───────────────┤
│ Research Agent│
│ Coding Agent │
│ Vision Agent │
│ Data Agent │
│ Review Agent │
└───────────────┘
↓
EXECUTION
↓
VERIFICATION
↓
HUMAN / POLICY GATE
↓
FINAL RESULT
End of Chapter 15
Comments
Post a Comment