🌈 FAISS search_kwargs 

When you convert a FAISS vector store into a retriever:

retriever = vectorstore.as_retriever(
    search_type="similarity",
    search_kwargs={"k": 5}
)

search_kwargs decides HOW the retriever searches your vector index.

Let’s break down all keyword arguments in a colorful and beautiful way 👇

🔵 1. k — Number of Results

Meaning:
How many similar documents do you want back?

Default: Usually 4

You use it when:
You want the top K most similar chunks.

Example:

{"k": 5}

🔍 “Bring me the top 5 closest matches!”

🟢 2. score_threshold — Minimum Similarity Needed

Used only in:
"similarity_score_threshold" or "similarity_distance_threshold" retrievers.

Meaning:
Return results ONLY when similarity score > threshold.

Example:

{
    "score_threshold": 0.8,
    "k": 20    # internal search depth
}

🛑 “Ignore weak matches. Give me only strong, high-score documents!”

🟣 3. fetch_k — Candidates Before Final Selection (MMR Only)

Used only in:
search_type="mmr" (Maximum Marginal Relevance)

Meaning:
Search this many candidates internally before selecting diverse ones.

Example:

{
    "k": 5,         # final 5 results
    "fetch_k": 25   # initial candidate pool
}

🔎 “Search 25 items first, then pick the 5 best diverse ones!”

🟠 4. lambda_mult — Diversity Factor (MMR Only)

Used only in:
search_type="mmr"

Meaning:
Controls trade-off between similarity and diversity.

Example:

{"lambda_mult": 0.5}

🎨 “Blend similarity and diversity 50-50!”

🟡 5. filters / filter — Metadata Filtering

(Some FAISS wrappers support this)

Meaning:
Filter only documents with matching metadata.

Example:

{
    "k": 4,
    "filters": {"source": "policies"}
}

📁 “Search only inside documents where source = policies!”

🔴 6. score_threshold (for Distance)

Used in:
search_type="similarity_distance_threshold"

Meaning:
Based on distance (lower distance = closer match).
Only return documents where distance < threshold.

💠 FAISS Search Types & Allowed kwargs

🌟 COMPLETE SUMMARY