Vector databases (Pinecone, ChromaDB, Weaviate) in Prompt Engineering / GenAI - Model Metrics & Evaluation

Metrics & Evaluation - Vector databases (Pinecone, ChromaDB, Weaviate)

Which metric matters for Vector Databases and WHY

Vector databases store and search data by similarity, not exact matches. The key metric is Recall, which tells us how many of the truly similar items the database finds. High recall means the database finds most relevant vectors, important for good search results.

Precision also matters because it shows how many found items are actually relevant. But recall is often more critical because missing relevant items hurts user experience more than extra irrelevant ones.

Another important metric is Latency -- how fast the database returns results. Fast responses keep users happy.

Confusion Matrix for Vector Search

    |---------------------------|
    |           | Predicted     |
    | Actual    | Similar | Not |
    |-----------|---------|-----|
    | Similar   | TP      | FN  |
    | Not       | FP      | TN  |
    |---------------------------|

    TP = Relevant vectors found
    FN = Relevant vectors missed
    FP = Irrelevant vectors found
    TN = Irrelevant vectors not found
    

Recall = TP / (TP + FN) shows how many relevant vectors were found.

Precision = TP / (TP + FP) shows how many found vectors are relevant.

Precision vs Recall Tradeoff with Examples

Imagine a vector database for a movie recommendation app.

• High Recall, Lower Precision: The database returns many movies similar to your favorite, including some less relevant ones. You get more options but some may not fit your taste.
• High Precision, Lower Recall: The database returns only very close matches. You get fewer options but they are very relevant.

For discovery, high recall is better to not miss good options. For strict filtering, high precision is better to avoid irrelevant results.

What Good vs Bad Metrics Look Like for Vector Databases

• Good: Recall above 90%, Precision above 80%, Latency under 100ms. Most relevant vectors found quickly.
• Bad: Recall below 50%, Precision below 50%, Latency over 500ms. Many relevant vectors missed or slow responses.

Good metrics mean users find what they want fast. Bad metrics mean poor search experience.

Common Pitfalls in Vector Database Metrics

• Ignoring Recall: Focusing only on precision can miss many relevant vectors.
• Data Leakage: Testing on vectors already in the database inflates metrics.
• Overfitting: Tuning only for a small test set can hurt real-world performance.
• Latency Overlooked: Fast search is critical; slow queries frustrate users.

Self Check

Your vector database returns results with 98% precision but only 12% recall. Is it good for production?

Answer: No. While most returned vectors are relevant (high precision), the database misses most relevant vectors (very low recall). Users will not see many good matches, hurting experience.

Key Result

Recall is key for vector databases to find most relevant items; precision and latency also matter for quality and speed.

Practice

1. What is the main purpose of a vector database like Pinecone, ChromaDB, or Weaviate?

easy

A. To store plain text documents only

B. To perform traditional SQL queries on structured data

C. To store and search data based on similarity using number lists

D. To create visual graphs from data

Solution

1. Step 1: Understand what vector databases store

   Vector databases store data as vectors, which are lists of numbers representing complex data like images or text.

2. Step 2: Identify the main use of vector databases

   They allow fast searching by similarity, not by exact matches like traditional databases.

3. Final Answer:

   To store and search data based on similarity using number lists -> Option C

4. Quick Check:

   Vector databases = similarity search [OK]

Hint: Vector DBs = search by meaning, not exact text [OK]

Common Mistakes:

• Thinking vector DBs only store text
• Confusing vector DBs with SQL databases
• Assuming vector DBs create visual graphs

2. Which of the following is the correct way to insert a vector into Pinecone using Python?

easy

A. pinecone.insert(id='vec1', vector=[0.1, 0.2, 0.3])

B. pinecone.upsert(vectors=[('vec1', [0.1, 0.2, 0.3])])

C. pinecone.add_vector('vec1', [0.1, 0.2, 0.3])

D. pinecone.push_vector(id='vec1', vector=[0.1, 0.2, 0.3])

Solution

1. Step 1: Recall Pinecone's method to add vectors

   Pinecone uses the 'upsert' method to insert or update vectors, which takes a list of tuples with id and vector.

2. Step 2: Match the correct syntax

   pinecone.upsert(vectors=[('vec1', [0.1, 0.2, 0.3])]) uses 'upsert' with a list of tuples, which is the correct syntax.

3. Final Answer:

   pinecone.upsert(vectors=[('vec1', [0.1, 0.2, 0.3])]) -> Option B

4. Quick Check:

   Use upsert with list of (id, vector) tuples [OK]

Hint: Pinecone uses upsert() with list of (id, vector) [OK]

Common Mistakes:

• Using insert() instead of upsert()
• Passing vector without wrapping in a list
• Using non-existent methods like add_vector or push_vector

3. Given the following code snippet using ChromaDB, what will be the output?

collection.add(ids=['1'], embeddings=[[0.1, 0.2, 0.3]], metadatas=[{'type': 'image'}], documents=['cat image'])
results = collection.query(query_embeddings=[[0.1, 0.2, 0.3]], n_results=1)
print(results['documents'])

medium

A. [['cat image']]

B. ['cat image']

C. [{'type': 'image'}]

D. []

Solution

1. Step 1: Understand what add() does in ChromaDB

   The add() method stores the document with its vector and metadata in the collection.

2. Step 2: Understand query() output format

   The query() method returns a dictionary with keys like 'documents' containing a list of lists of matched documents.

3. Final Answer:

   [['cat image']] -> Option A

4. Quick Check:

   Query returns list of lists of documents [OK]

Hint: ChromaDB query returns list of lists for documents [OK]

Common Mistakes:

• Expecting a flat list instead of list of lists
• Confusing documents with metadata
• Assuming empty result when vector matches exactly

4. You wrote this Weaviate query to find similar items but get an error:

client.query.get('Article', ['title']).with_near_vector({'vector': [0.1, 0.2]}).do()

What is the likely cause of the error?

medium

A. The query must include a filter parameter

B. The method with_near_vector does not exist in Weaviate client

C. The class name 'Article' must be lowercase

D. The vector length is too short; it should match the database dimension

Solution

1. Step 1: Check vector length requirement in Weaviate

   Weaviate expects the vector length to match the dimension used when creating the index, usually 3 or more numbers.

2. Step 2: Identify the error cause

   The vector [0.1, 0.2] has length 2, which is likely shorter than expected, causing the error.

3. Final Answer:

   The vector length is too short; it should match the database dimension -> Option D

4. Quick Check:

   Vector length must match index dimension [OK]

Hint: Vector length must match index dimension in Weaviate [OK]

Common Mistakes:

• Thinking method name is wrong
• Assuming class names must be lowercase
• Believing filter is always required

5. You want to build a search system that finds similar product descriptions using Weaviate. Which steps should you follow to prepare and query the data correctly?

hard

A. Create a schema with a vector index, add product descriptions as objects with vectors, then query using nearVector filter

B. Store product descriptions as plain text only, then query with SQL-like text search

C. Upload product images only, then query using image metadata filters

D. Create a schema without vector index, add descriptions, then query using exact match filters

Solution

1. Step 1: Define schema with vector index in Weaviate

   To search by similarity, the schema must include a vector index for the product description class.

2. Step 2: Add product descriptions as objects with vectors

   Each product description is stored as an object with its vector embedding representing meaning.

3. Step 3: Query using nearVector filter

   Use the nearVector filter in queries to find objects with vectors close to the query vector.

4. Final Answer:

   Create a schema with a vector index, add product descriptions as objects with vectors, then query using nearVector filter -> Option A

5. Quick Check:

   Schema + vectors + nearVector query = correct approach [OK]

Hint: Schema with vectors + nearVector query = similarity search [OK]

Common Mistakes:

• Trying to search plain text without vectors
• Using exact match filters for similarity search
• Ignoring schema vector index setup