Text chunking strategies in Prompt Engineering / GenAI - Model Pipeline Trace

Model Pipeline - Text chunking strategies

This pipeline breaks long text into smaller, manageable pieces called chunks. These chunks help AI models understand and process text better by focusing on smaller parts at a time.

Data Flow - 5 Stages

1Input Text

1 document x 1000 words→Receive raw text document→1 document x 1000 words

"Once upon a time in a faraway land, there was a small village surrounded by mountains..."

↓

2Preprocessing

1 document x 1000 words→Clean text (remove punctuation, lowercase)→1 document x 980 words

"once upon a time in a faraway land there was a small village surrounded by mountains"

↓

3Chunking

1 document x 980 words→Split text into chunks of 100 words with 20 words overlap→12 chunks x 100 words each

Chunk 1: words 1-100, Chunk 2: words 81-180, ..., Chunk 12: words 881-980

↓

4Feature Extraction

12 chunks x 100 words→Convert each chunk into numerical vectors (embeddings)→12 chunks x 100-dimensional vectors

[0.12, 0.45, ..., 0.33] for chunk 1

↓

5Model Input

12 chunks x 100-dimensional vectors→Feed chunks into AI model for understanding or generation→Model processes each chunk independently

Model predicts sentiment or answers questions per chunk

Training Trace - Epoch by Epoch

Loss
1.0 |****
0.8 |*** 
0.6 |**  
0.4 |*   
0.2 |    
0.0 +----
     1 2 3 4 5 Epochs

Epoch	Loss ↓	Accuracy ↑	Observation
1	0.85	0.60	Model starts learning from chunked text with moderate accuracy
2	0.65	0.72	Loss decreases and accuracy improves as model adapts to chunked inputs
3	0.50	0.80	Model shows good understanding of chunks, accuracy rising
4	0.40	0.85	Training converges with lower loss and higher accuracy
5	0.35	0.88	Final epoch shows stable performance on chunked text

Prediction Trace - 4 Layers

Layer 1: Input Chunk

Layer 2: Embedding Layer

Layer 3: Model Processing

Layer 4: Output Aggregation

Model Quiz - 3 Questions

Why do we use overlapping words between chunks?

A To keep context between chunks

B To reduce the total number of chunks

C To make chunks shorter

D To remove stop words

Key Insight

Breaking long text into overlapping chunks helps AI models keep context and understand text better. Training shows that chunking improves learning by making text easier to process in parts.

Practice

1. What is the main purpose of text chunking in AI models?

easy

A. To generate new text from scratch

B. To split long text into smaller, manageable pieces

C. To remove stop words from text

D. To translate text into different languages

Solution

1. Step 1: Understand the concept of text chunking

   Text chunking means breaking a long text into smaller parts so it is easier to handle.

2. Step 2: Identify the main goal in AI context

   This helps AI models process and understand large texts better by working on smaller pieces.

3. Final Answer:

   To split long text into smaller, manageable pieces -> Option B

4. Quick Check:

   Text chunking = splitting text [OK]

Hint: Chunking means breaking text into smaller parts [OK]

Common Mistakes:

• Confusing chunking with translation
• Thinking chunking removes words
• Believing chunking generates new text

2. Which of the following is a correct way to create overlapping text chunks in Python?

easy

A. chunks = [text[i:i+chunk_size] for i in range(0, len(text), overlap)]

B. chunks = [text[i:i+chunk_size] for i in range(0, len(text), chunk_size)]

C. chunks = [text[i:i+chunk_size] for i in range(0, len(text), chunk_size - overlap)]

D. chunks = [text[i:i+chunk_size] for i in range(overlap, len(text), chunk_size)]

Solution

1. Step 1: Understand overlapping chunk logic

   To create overlapping chunks, the step size must be smaller than chunk size by the overlap amount.

2. Step 2: Check the range step in options

   chunks = [text[i:i+chunk_size] for i in range(0, len(text), chunk_size - overlap)] uses chunk_size - overlap as step, correctly creating overlaps.

3. Final Answer:

   chunks = [text[i:i+chunk_size] for i in range(0, len(text), chunk_size - overlap)] -> Option C

4. Quick Check:

   Overlap step = chunk_size - overlap [OK]

Hint: Overlap step = chunk size minus overlap length [OK]

Common Mistakes:

• Using chunk_size as step (no overlap)
• Using overlap as step (too small steps)
• Starting range at overlap instead of zero

3. Given text = 'abcdefghij', chunk_size = 4, and overlap = 2, what is the output of this code?

chunks = [text[i:i+chunk_size] for i in range(0, len(text)-overlap, chunk_size - overlap)]
print(chunks)

medium

A. ['abcd', 'cdef', 'efgh', 'ghij']

B. ['abcd', 'efgh', 'ij']

C. ['abcd', 'bcde', 'cdef', 'defg']

D. ['abcd', 'bcdf', 'cdeg', 'defh']

Solution

1. Step 1: Calculate step size

   Step = chunk_size - overlap = 4 - 2 = 2.

2. Step 2: Generate chunks using step 2

   Chunks are:
   i=0: text[0:4] = 'abcd'
   i=2: text[2:6] = 'cdef'
   i=4: text[4:8] = 'efgh'
   i=6: text[6:10] = 'ghij'

3. Final Answer:

   ['abcd', 'cdef', 'efgh', 'ghij'] -> Option A

4. Quick Check:

   Chunks overlap by 2 chars = ['abcd', 'cdef', 'efgh', 'ghij'] [OK]

Hint: Step = chunk size minus overlap; slice text accordingly [OK]

Common Mistakes:

• Ignoring overlap and stepping by chunk size
• Wrong slicing indices
• Confusing overlap with chunk size

4. This code aims to chunk text with overlap but has a bug:

chunk_size = 5
overlap = 2
chunks = []
for i in range(0, len(text), chunk_size + overlap):
    chunks.append(text[i:i+chunk_size])
print(chunks)

What is the error?

medium

A. Step size should be chunk_size - overlap, not chunk_size + overlap

B. Chunk size should be increased by overlap

C. Overlap should be zero for chunking

D. The loop should start at overlap, not zero

Solution

1. Step 1: Understand step size for overlapping chunks

   To create overlap, step size must be less than chunk size by overlap amount.

2. Step 2: Identify incorrect step in code

   Code uses chunk_size + overlap which skips overlap, causing gaps.

3. Final Answer:

   Step size should be chunk_size - overlap, not chunk_size + overlap -> Option A

4. Quick Check:

   Overlap step = chunk_size - overlap [OK]

Hint: Overlap step = chunk size minus overlap, not plus [OK]

Common Mistakes:

• Adding overlap instead of subtracting
• Setting overlap to zero incorrectly
• Changing loop start index wrongly

5. You have a very long document and want to chunk it for an AI model. You want each chunk to have 100 words and overlap by 20 words to keep context. Which strategy balances chunk size and context best?

hard

A. Use chunk size 80 and step size 100 to create non-overlapping chunks

B. Use chunk size 100 and step size 100 to create overlapping chunks

C. Use chunk size 120 and step size 100 to create overlapping chunks

D. Use chunk size 100 and step size 80 (100 - 20) to create overlapping chunks

Solution

1. Step 1: Define chunk and step sizes for overlap

   Chunk size is 100 words, overlap is 20 words, so step size = 100 - 20 = 80.

2. Step 2: Choose correct step size to maintain overlap

   Step size 80 means each chunk starts 80 words after previous, overlapping 20 words.

3. Final Answer:

   Use chunk size 100 and step size 80 (100 - 20) to create overlapping chunks -> Option D

4. Quick Check:

   Step = chunk size - overlap = 80 [OK]

Hint: Step size = chunk size minus overlap for best context [OK]

Common Mistakes:

• Using step size larger than chunk size
• Setting overlap to zero accidentally
• Confusing chunk size with step size