Rerank

Takes a search query and a list of candidate documents, returns a stronger ranking of which document best answers the query. The second pass after embedding-based top-K retrieval.

Available models

Model	Type
verirerag (recommended)	TEI (Text Embeddings Inference) cross-encoder, multilingual
Cohere rerank-multilingual-v3.0	If a Cohere-keyed channel exists

Why rerank?

Embedding similarity is fast (vector dot product) but "closest neighbour" isn't always "correct answer". A cross-encoder examines each (query, document) pair and scores actual relevance. It's slower — so we first cut down to 50–100 with embedding, then verirerag picks the real top 3.

[5000 docs via embedding → top 50]  →  [verirerag → top 5]

curl

curl https://app.qevron.ai/v1/rerank \
  -H "Authorization: Bearer sk-..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "verirerag",
    "query": "what is qevron",
    "documents": [
      "The weather is nice today.",
      "Qevron is an AI gateway application.",
      "Channels follow the OpenAI schema.",
      "Turkish and English are supported."
    ]
  }'

Response

{
  "results": [
    {"index": 1, "relevance_score": 0.96, "document": {"text": "Qevron is an AI gateway application."}},
    {"index": 2, "relevance_score": 0.62, "document": {"text": "Channels follow the OpenAI schema."}},
    {"index": 3, "relevance_score": 0.18, "document": {"text": "Turkish and English are supported."}},
    {"index": 0, "relevance_score": 0.02, "document": {"text": "The weather is nice today."}}
  ]
}

index is the position in the original documents array — use it to look up your own metadata.

Python (raw requests — OpenAI SDK doesn't expose rerank)

import requests

resp = requests.post(
    "https://app.qevron.ai/v1/rerank",
    headers={"Authorization": "Bearer sk-..."},
    json={
        "model": "verirerag",
        "query": "what is qevron",
        "documents": ["d1", "d2", "d3"],
    },
)
results = resp.json()["results"]
# already sorted by relevance_score desc

In a RAG pipeline

# 1) Embedding picks candidates
candidates = vector_db.search(embed(user_query), top_k=50)

# 2) Rerank computes real relevance
reranked = rerank("verirerag", user_query, [c.text for c in candidates])

# 3) Top 3 go into the LLM as context
top_3 = [candidates[r["index"]] for r in reranked[:3]]

Troubleshooting

• All scores very low (< 0.05) → query likely doesn't overlap with documents. Audit embedding-side candidate selection first.
• Slow → 100+ documents are heavy. Trim with embedding to ~50 before rerank.
• OpenAI SDK doesn't work → correct, rerank isn't in OpenAI's standard; use raw requests/httpx.

Next: Embedding · RAG walkthrough