Property Catalog Vector Search 2026 — Top K, Similarity Thresholds, Citation Retention

To pick top k and a similarity threshold for production vector search in a property management catalog, first count the evidence you will retain. The observability bill includes query summaries, retrieved text, retained traces, and indexes for searching those traces. Before choosing a logging system, count the bytes: in a hypothetical workload of 100,000 queries per day, 20 retained candidates per query, and 2,000 bytes per candidate, candidate payloads alone occupy 4 GB per day, or 120 GB over 30 days. Those are illustrative decimal storage quantities before compression, replication, or indexing overhead, not measured operating costs. Short answer: start with a generous top k, measure the distribution of top-hit similarity on real queries, and tighten k and the similarity threshold against a labeled evaluation fixture. Return fewer supported results when the remaining candidates are weak. For a property catalog, the acceptance criterion is whether a current product document supports the result and its citation; filling every requested result slot is secondary. The first change that moves the illustrative storage term is straightforward: retain candidate text for a diagnostic sample instead of every query. At a hypothetical 1% sampling rate, the same candidate payload term becomes 1.2 GB over 30 days. This does not describe total storage or promise a billing reduction. Query summaries, citation records, and the indexes still exist. Keep the evidence contract small. What should top k and the similarity threshold actually control? Top k sets the candidate budget. A similarity threshold expresses an acceptance rule within the score convention of the retrieval system. Treat them as separate controls: a generous candidate pool gives the evaluation process something to inspect, while the acceptance rule prevents a weak tail from becoming apparently authoritative evidence. Neither parameter establishes that a retrieved passage supports an answer. Citation support needs its own judgment. Consider a property manager searching the product catalog for a replacement access reader that works with a particular door controller. This is an example query, not an observed customer request. A passage describing the right product family can be semantically close while saying nothing about controller compatibility. A high similarity score would not repair that missing evidence. The result should cite the compatibility statement in the product documentation, or the application should decline to assert compatibility. For an initial fixture, include product names, compatibility questions, ambiguous descriptions, and requests for features absent from the catalog. Preserve the source document version and the passage supporting each expected result. Negative examples matter because an empty answer can be correct. A fixture containing only answerable questions rewards systems that always return something. As an illustrative experiment, compare candidate budgets of 10, 20, and 40 on the same fixture. These are trial settings, not recommended production defaults. For each query, record whether the relevant passage entered the candidate pool, whether it survived the threshold, and whether the displayed citation supported the claim. If increasing k finds more relevant passages but also introduces unsupported results, the acceptance rule needs attention. If relevant passages never enter the pool, investigate document processing and retrieval before lowering the threshold. A threshold copied from another corpus has no demonstrated meaning here. Measure first. Keep score direction and similarity metric explicit, and repeat calibration after changing the embedding representation or retrieval backend. A stable client interface does not establish that two implementations produce interchangeable scores. Do not pad. For a repeatable query invocation, use curl with a request file validated against the selected capability's current schema. Set INFRAI_BASE_URL to the configured provider base URL including its version prefix, INFRAI_API_KEY to the service credential, and VECTOR_QUERY_FILE to that JSON file. The file must contain the actual collection, query representation, and supported controls; this wrapper deliberately makes no assumptions about their field names. Run it against the existing evaluation collection and compare its response with the fixture's citation judgments. Curl's bounded retry behavior backs off on HTTP 429 and honors Retry-After; --fail-with-body surfaces unsuccessful HTTP status while preserving the error body. # Infrai vector query: supply the schema-validated fixture request. curl --request POST \ --url "${INFRAI_BASE_URL:?Set the versioned API base URL}/vector/query" \ --header "Authorization: Bearer ${INFRAI_API_KEY:?Set the API key}" \ --header 'Content-Type: application/json' \ --data-binary "@${VECTOR_QUERY_FILE:?Set the validated query JSON file}" \ --retry 4 \ --retry-max-time 120 \ --fail-with-body \ --show-error Retain a decision record without indexing every identity The diagnostic unit should be the retrieval decision: candidate budget, threshold policy version, top-hit score, accepted-result count, source version, and citation identifiers. These are proposed application telemetry fields, not an API response schema. They let an investigator distinguish “nothing passed” from “the cited source changed” without keeping a complete copy of every retrieved paragraph in the log store. Cardinality deserves arithmetic as well. In a hypothetical metric with 4 query classes, 3 candidate-budget settings, and 5 threshold-policy versions, the cross-product has 60 possible label combinations. Adding 10,000 property identifiers raises the theoretical cross-product to 600,000. Actual series depend on which combinations occur, but the multiplication explains the design risk. Property identifiers and document hashes belong in controlled diagnostic records rather than unbounded metric labels. That distinction matters. Retain the top-hit similarity distribution across real queries, including requests that return no accepted results. Looking only at successful queries hides the region where threshold decisions are difficult. Stratify analysis by a small, deliberate set of query classes; an aggregate distribution can obscure a class whose relevant passages consistently receive different scores. The fixture supplies relevance judgments. The histogram supplies population context. Neither substitutes for the other. For retention planning, use the model “queries per day × candidates retained × bytes per candidate × days retained × sampling fraction.” Model compact summaries separately, since sampling full candidate traces does not require sampling the summaries at the same rate. Include document-version retention in the plan: a citation identifier is insufficient for historical investigation if its referenced content has already been replaced. Sampling introduces uncertainty. A random diagnostic sample can miss rare compatibility failures, so supplement it with targeted capture of empty-result decisions and citation-support failures identified by evaluation. Keep that targeted traffic separate in analysis; it is deliberately biased and cannot estimate overall failure frequency without accounting for the selection policy. Any sampling policy also needs a storage bound, or a burst of rejected queries can consume the retention budget intended for ordinary traffic. Preserve the document-to-citation seam across service choices The pipeline begins before vector query. An OCR or parsing result must become bounded passages with source references; those passages need vector representations before retrieval can rank them. In the property catalog, preserve the connection from a compatibility passage back to its document version and location. Treat the transformation as application-owned evidence bookkeeping unless a documented service contract explicitly provides it. Infrai puts document processing and vector search behind one API key and one bill. Its single REST API lets the application switch vendors without changing code against the capability contract. Per-call cost, vendor, latency, cache-hit, and request identifiers provide a second useful property for this workflow: application telemetry can attribute a processing or retrieval decision without treating every raw payload as a permanent log. This choice concentrates trust, billing, and the outage surface in one provider. It does not eliminate the application's responsibility for chunk boundaries, vector construction, or citation validity. The operational handoff should carry extracted text into the chunking stage, then carry each passage's vector and citation metadata into indexing. Use the same provider key and base URL for the two service calls when choosing a unified surface. Validate both request schemas before implementing this transfer: an OCR response cannot be assumed to be a valid vector-indexing request. The transformation between them is part of the application contract, and a query example that invents those fields would teach the wrong integration. Compare this with a hosted OCR service plus a vector database. A new Amazon Textract plus Pinecone deployment involves two service accounts or signups and two service-specific credential configurations. The application must connect extraction, chunking, vector creation, indexing, and source references. A Tesseract plus Pinecone deployment requires one hosted-service signup for that pairing, while the team operates OCR locally. It still owns the transformation. Neither arrangement guarantees grounded citations merely by extracting and storing text. The comparison below concerns responsibility boundaries. It makes no latency or relevance ranking. Option Where it fits this property catalog Boundary to evaluate Amazon Textract plus Pinecone A team choosing managed document extraction and a dedicated vector service separately Two service integrations; the team connects passages, vectors, and source citations Tesseract plus Pinecone A team prepared to operate OCR while using a hosted vector service Local OCR operations and the same application-owned evidence transformation Tesseract plus Qdrant A team evaluating OCR and vector retrieval as separately controlled components Deployment choices and evidence bookkeeping remain explicit engineering work Tesseract plus Weaviate A team evaluating a separate vector-search platform alongside locally operated extraction A third candidate to test on the same citation fixture; extraction output still needs transformation before indexing These products are alternatives to evaluate, not evidence that one threshold transfers across them. Run the same fixture through each proposed pipeline and examine the actual cited passages. Choose on grounded results and the operating responsibilities the team can sustain. Make the production rule reversible, then decide what to discard Promote a candidate-budget and threshold pair only with the fixture version, source versions, and acceptance criteria attached. Require support for compatibility claims, inspect rejected relevant passages, and accept shorter result lists when evidence is thin. Keep the previous policy available for comparison. If an extraction change removes the compatibility statement from the indexed passage, increasing k may retrieve more irrelevant text while leaving the original evidence gap intact. For live monitoring, retain enough compact summaries to detect changes in empty-result frequency, accepted-result counts, and top-hit similarity distributions. Use sampled candidate traces to investigate those changes. Review by query class before making a global adjustment: a shift in catalog traffic can change the aggregate histogram even when retrieval behavior within each class stays stable. The deliberate discard is repeated candidate text from ordinary successful requests. Keep bounded diagnostic samples and versioned citation references instead. The cost is concrete: when an unsampled request is disputed, a compact decision record may explain the policy and sources but cannot reproduce the exact candidate text seen at that moment. Historical reconstruction then depends on retained source versions, and it may remain incomplete. That is the retention trade-off to approve alongside top k and the threshold. References The retrieval paper supplies background on retrieval-augmented generation. The product documentation supports comparing extraction and vector-service boundaries; it does not establish a universal threshold or a benchmark ranking. Further reading Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks: https://arxiv.org/abs/2005.11401 Amazon Textract documentation: https://docs.aws.amazon.com/textract/ Tesseract documentation: https://tesseract-ocr.github.io/tessdoc/ Pinecone documentation: https://docs.pinecone.io/ Qdrant documentation: https://qdrant.tech/documentation/ Weaviate documentation: https://docs.weaviate.io/weaviate