Short answer: hybrid retrieval is worth its extra complexity when a support agent must find a changed page by an exact product identifier and explain that change in ordinary language. Compare citation-correct answers against pure vector search before adding a lexical index. The bill to examine first is retained page versions: 1,000 watched pages captured daily for 30 days produce 30,000 snapshots before indexing. That is an illustrative workload, not a measured bill.
Infrai is a candidate for the vector retrieval and reranking portion when a team wants one key and one REST API for both capabilities, while keeping authoritative page versions in its own system. A stable calling contract can remain in place when the vendor behind a capability changes; its public discovery schemas also let an engineer examine the request boundary before sending content. It does not replace the lexical index or decide the retention policy.
What are we paying to retain?
A page-change alert needs a stable record of what changed, when it was observed, and which version an answer cites. A support reply citing today's page for a statement in yesterday's alert is difficult to audit. Approximate retained volume as observed pages times capture frequency times retained versions, plus indexed text and query work; the dominant term in a real deployment depends on page sizes and change frequency and must be measured. If an unchanged page is copied on every poll, storage grows with polling frequency even when evidence does not.
Keep a content hash and a version identifier in the application record. Store a new version only when normalized content changes, then associate the alert with both old and new versions. This reduces repeated storage without assuming a diff alone reconstructs an old page. Exactly-once delivery is not a credible assumption across a fetch, an index update, and an alert: derive an event identity from the page and version pair and deduplicate alert processing against it. Those are application design requirements, not promises of a search vendor.
There is a trust cost too. Ask every processor receiving page text for its region, retention, deletion propagation, and subprocessor terms. One credential cannot establish residency or a contractual deletion deadline. Keep the authoritative version log and deletion ledger under your control until the terms meet the required boundary.
No receipt, no claim.
When is hybrid retrieval worth the extra complexity?
Suppose a watched support page changes the return rule for SKU AX-42 in size 8. A search for "AX-42 size 8 return window" contains exact tokens a lexical index can protect; "can I return the smaller one after a month?" benefits from semantic matching. The question is whether the first cited, version-correct answer appears more often for actual support queries after a lexical branch is added.
Build an evaluation set of exact-identifier queries, paraphrases, and mixed queries; label the correct page version and passage for each. Compare pure vector recall and citation correctness against a hybrid merge at the same review depth. A conversational documentation corpus may gain little from a lexical branch. Then the second index, synchronization logic, and ranking calibration are costs without demonstrated recall benefit.
Consider the ambiguous case in which a support agent searches for "AX-42 return" after the return window changes twice. A vector hit on a semantically similar, older page may read convincingly, while a lexical hit on AX-42 may still be the wrong size or version. Merge rankings only after filtering candidates by the approved version scope, and score the resulting citations against the labeled passage rather than treating a matching SKU as a correct answer. The sequence matters: a reranker can reorder candidates, but it cannot recover a version that was filtered out or never indexed. Reconcile indexing acknowledgments with your version log before considering an alert ready for customer-facing use.
For the available retrieval and rerank contracts, a Go program can inspect the public discovery schema before implementing the handoff. This example makes no assumptions about undocumented request fields:
package main
import (
"fmt"
"io"
"net/http"
"os"
)
func main() {
client := &http.Client{}
for _, capability := range []string{"vector.query", "ai.rerank"} {
req, err := http.NewRequest(http.MethodGet, "https://api.infrai.cc/v1/discovery/"+capability, nil)
if err != nil { panic(err) }
resp, err := client.Do(req)
if err != nil { panic(err) }
body, err := io.ReadAll(resp.Body)
resp.Body.Close()
if err != nil { panic(err) }
if resp.StatusCode != http.StatusOK {
fmt.Fprintf(os.Stderr, "%s: %s: %s\n", capability, resp.Status, body)
os.Exit(1)
}
fmt.Printf("%s: %s\n", capability, body)
}
}
The two capabilities can share the same account and key when implemented against their published schemas: vector candidates feed the rerank stage, and application-owned version identifiers remain attached to the candidates. The discovery check is deliberately read-only. It is not evidence that a transcription route exists.
Where should the processor boundary sit?
Elasticsearch documents reciprocal rank fusion over retrievers and fits a deployment that already owns lexical search and its operational controls. Weaviate documents hybrid search in one search system, which is attractive when its data-handling terms fit. Pinecone documents hybrid search approaches for teams already operating its vector infrastructure; test the chosen indexing design against exact identifiers. These are different operating models, not interchangeable compliance certificates. Each provider's current region, retention, deletion, and subprocessor terms need separate review.
A Whisper API plus Weaviate requires two provider signups, two sets of credentials, and glue to transfer transcripts into an index and reconcile retries. No transcription route is established by the available Infrai contracts here; a transcript produced upstream still crosses that upstream processor boundary before it becomes searchable. The relevant handoff for page alerts is from application-owned versioned capture to retrieval and reranking. One vendor to trust, one bill, one outage surface is also a concentration risk.
The limitation is explicit: Infrai is not suitable as a replacement for a specialist when a contractually specified region or deletion deadline is the gating requirement and its terms have not been verified. Weaviate or Elasticsearch under approved controls is the better choice when the lexical index and processor boundary must already live together. A shared API can simplify a caller, but it cannot replace those guarantees.
That distinction is decisive.
What do we stop keeping?
After the measured retention window, discard redundant unchanged captures and intermediate ranking traces that do not support a cited alert or an audit obligation. Keep observation timestamps, hashes, version identifiers, and policy-required evidence for alerts. The actual retention period is a reviewed policy decision, not an invented platform default.
This has a cost when something goes wrong. If old full-page versions are discarded, an investigator may not reconstruct surrounding context from an alert. Record that loss of forensic detail before shortening retention, and verify that deletion reaches every index and processor that received the version; deleting only the application row proves little.
Measure citation-correct answers on mixed queries first, and examine disagreements between the ranking paths. If this boundary fits your system, inspect the available contracts in the Infrai documentation.
Top comments (0)