AI-Powered Enterprise Search with Hybrid Ranking and Permissions
AI enterprise search is a search experience over internal content that combines keyword matching with vector (semantic) retrieval, machine-learned re-ranking and query understanding, while showing each user only documents they are permitted to open. NTSPL builds and tunes these search platforms for firms whose staff lose hours locating the right file, drawing, manual or precedent.
Delivered on
In scope
What Makes Search Intelligent: Query Understanding, Hybrid Retrieval, Re-ranking
Ask an associate, an engineer or a service technician how they find documents and the answer is usually a colleague, not the intranet. SharePoint search ranks by recency and title words, the DMS search needs exact matter numbers, and file-share search does not exist. Typing 'clutch judder after rain' returns nothing because the bulletin says 'take-up shudder under high humidity'. Searches with part numbers return 400 hits sorted alphabetically. Every system has its own box, so people search four places and still miss the document that matters.
NTSPL builds one search layer across those sources, with relevance treated as a measurable engineering target. We collect real query logs, build judgement lists with subject experts, and tune hybrid keyword and vector ranking, synonyms, field boosts and a re-ranker against NDCG and success-at-one. Facets come from document metadata that we enrich where it is missing. Permissions are mirrored from each source so that results respect matter walls and dealer boundaries. AI answers are optional, layered on top only once the ranking is right.
Symptoms of a Search Problem Worth Fixing
Signals you need Enterprise Search with AI
- Staff report spending more than 30 minutes a day looking for documents, or a survey shows under 40% find what they need on the first search.
- Content sits in 3 or more repositories, such as SharePoint, a DMS, file shares and a portal, each with its own search box.
- Zero-result rates exceed 15% because users describe problems in words the documents do not use.
- Part numbers, clause references or matter codes must match exactly while descriptive queries need semantic matching.
- Ethical walls, dealer boundaries or project confidentiality mean search results must differ by user.
When to choose something else
- If users want a composed answer with citations rather than a ranked list, build a RAG knowledge assistant, ideally on top of the search index described here.
- If content is fewer than about 5,000 documents in one well-organised library, native SharePoint or DMS search with better metadata may be enough.
- If the real problem is that documents do not exist or are out of date, fix content ownership first; better search will surface the gap faster but not close it.
Search Platform Components NTSPL Designs and Tunes
Query log and intent analysis
Analysis of 3-6 months of search logs to find top intents, zero-result queries, abandoned searches and vocabulary gaps.
Connectors and permission mirroring
Connectors to each repository with ACL capture, incremental sync and deletion handling, including matter walls and group nesting.
Index schema and enrichment
Field design, entity extraction for part numbers, clauses and parties, language detection and auto-tagging to populate facets.
Relevance engineering
Hybrid keyword and vector weights, field boosts, synonym maps, freshness decay and a re-ranker, tuned against expert judgement lists.
Search experience
Results page with facets, previews with highlighted passages, did-you-mean, saved searches and optional AI answer summary.
Relevance operations
Dashboards for zero-result and low-click queries, a monthly tuning cycle and a synonym governance process.
Federated Hybrid Search with Security Trimming
Federated Hybrid Search Index with Mirrored Permissions
Connectors pull content and ACLs from each repository into an enriched index holding text, vectors and facet fields. Queries are analysed, run as hybrid keyword and vector searches filtered by the user's groups, re-ranked and returned with facets and highlighted passages.
Select any component
Federated Hybrid Search Index with Mirrored Permissions: click any of its 10 components to see its role and the design choice NTSPL makes for it, or trace the 7-step primary path.
Azure AI Search, Amazon OpenSearch and Agent Search in Practice
OpenSearch hybrid queries with the neural search plugin and a normalisation processor give the most control over ranking; Amazon Kendra closed to new customers on 30 July 2026, so for managed connectors with less tuning we use Bedrock Knowledge Bases over OpenSearch.
Azure AI Search combines BM25, vector fields and Reciprocal Rank Fusion with a semantic ranker; skillsets enrich content during indexing and security filters use Entra group IDs.
Agent Search provides managed connectors, ACL-aware results and tunable ranking for unstructured and structured data; we add serving controls such as boosts and synonyms, and the Ranking API for custom pipelines.
Ranking, Synonym, Facet and Federation Decisions
Keyword, vector or hybrid
For enterprise content we almost always run hybrid. Vector search handles paraphrase; keyword search handles codes such as 2M51-0034 or Section 138. We detect code-like queries and weight keyword scoring higher for them. Hybrid typically improves NDCG@10 by 8-20% over either method alone on our judgement lists; if it does not on yours, the embedding model or chunk size is usually wrong.
Synonyms and domain vocabulary
Synonym maps fix the most zero-result queries for the least effort. We mine reformulations from search logs, where users search one phrase then another and click, and propose candidates to a named vocabulary owner. Keep synonym sets under about 2,000 entries and review quarterly; uncontrolled synonym lists create more irrelevant results than they fix.
Federated query or central index
A central index gives consistent ranking, facets and speed, but means copying content and ACLs. Federated queries leave content in place but merge results whose scores are not comparable. We choose a central index for the 2-4 highest-value sources and federate only long-tail sources, labelling federated results clearly in the UI.
Whether to add AI answers to search
Add a generated summary only when search relevance is already good, with success-at-one above about 60%, otherwise the summary amplifies bad results. Show it above the result list, cite the top passages, and suppress it when the re-ranker score is low. Many intents, such as finding a form or drawing, need a link, not a summary.
Query-Log-Driven Search Delivery in Six Steps
Query log mining
Extract top 500 queries, zero-result queries and reformulation chains from existing search logs and interviews.
Judgement list creation
Subject experts grade results for 150-250 representative queries on a 0-3 scale to create the relevance baseline.
Connectors, ACLs and schema
Configure connectors, permission mirroring, index schema and enrichment, then run a permission leak test.
Relevance tuning
Iterate hybrid weights, boosts, synonyms and re-ranking until NDCG@10 and success-at-one meet target.
Search UI and pilot
Launch the results page to 100-300 users and track clicks, reformulations and satisfaction.
Relevance operations hand-over
Hand over analytics dashboards, tuning runbook and synonym governance to the search owner.
What you receive
Relevance and Findability Measures: NDCG, Success-at-One, Zero Results
| KPI | Typical target |
|---|---|
| NDCG@10 on expert judgement list | 0.75 or higher, with 15-30% uplift over baseline |
| Success at first result for known-item queries | 65-80% |
| Zero-result query rate | Under 5% |
| Query reformulation rate | Reduced by a third from baseline |
| Search latency p95 | Under 800 ms without AI summary |
Search Projects That Disappoint, and Why
| Risk | How NTSPL prevents it |
|---|---|
| Search reveals documents behind ethical walls or dealer boundaries | Item-level ACL mirroring, nested-group expansion and a leak test with accounts from each restricted group before launch. |
| Relevance tuned by opinion drifts after launch | Judgement lists act as a regression suite; every ranking change must hold or improve NDCG before release. |
| Poor metadata leaves facets empty or misleading | Enrichment fills key facets automatically, and owners see a monthly metadata completeness score per source. |
| Users keep old habits and bypass the new search | Embed the search box in the tools people already use, and publish before-and-after findability numbers. |
Enterprise search vs a RAG chatbot: which should we build first?
Build search first in most cases. A RAG assistant is only as good as the retrieval underneath it, and search is that retrieval with a ranked list people can inspect. Search also serves intents a chatbot handles badly: finding a specific drawing, a signed version, a form or every document mentioning a party. Once success-at-one is above about 60% and permissions are proven, add a generated answer panel or a separate assistant on the same index. Going straight to a chatbot hides poor retrieval behind fluent text, and teams then tune prompts to fix what is really a ranking problem.
Key facts: Enterprise Search with AI
Enterprise Search with AI in Practice
2 Enterprise Search with AI scenarios, from Legal services (corporate law firm, M&A due diligence) and Automotive OEM (after-sales service), plus related platform case studies.
Law firm's due diligence teams find red-flag clauses across data rooms 3x faster
Auto OEM technicians find the right repair procedure on the first search 74% of the time
Infrastructure company's legal team reviews contracts in hours, not days
Agri-inputs company sees dealer secondary sales daily instead of 45 days late
Enterprise Search with AI: Frequently Asked Questions
How is AI search different from SharePoint or DMS search?
It adds semantic matching, so 'notice period for resignation' finds the clause titled 'termination by employee', and it ranks with a trained re-ranker rather than title words and dates. It also spans multiple repositories in one index, with permissions mirrored from each, and extracts metadata so users can filter by model, matter, client or date.
What does an enterprise search implementation cost?
A search layer over two to four repositories typically takes 10-12 weeks with indicative NTSPL fees of ₹16-32 lakh. Run cost depends on index size and query volume; for one to five million passages and a few thousand users, expect roughly ₹60,000 to ₹2.5 lakh a month for the search service and embeddings.
Can search respect matter walls and project confidentiality?
Yes. Each document's access list, including nested groups and ethical-wall exclusions, is captured from the source and stored with the index entry. Every query carries the user's resolved groups as a filter, so a restricted document never appears in results, counts or facets. We verify this with test accounts before go-live.
How do you measure whether search has improved?
We use judgement lists, where experts grade results for real queries, and compute NDCG@10 and success at first result before and after. In production we track zero-result rates, reformulations and clicks on the first three results. Satisfaction surveys alone are too noisy to steer tuning.
Can the search index stay in India?
Yes. Azure AI Search, Amazon OpenSearch Service and Google Cloud search services can be deployed in Indian regions such as Central India, Mumbai or Delhi. We confirm the embedding model also runs in-region, since sending passages to an embedding endpoint elsewhere would move data outside India.
Does the search work for Hindi or mixed-language queries?
Largely, yes. Multilingual embedding models handle Hindi and Hinglish queries reasonably for semantic matching, and we add transliteration synonyms for common terms. We test on your actual query logs, because performance on technical vocabulary in Indian languages varies by model and by how users type.
Scoping Enterprise Search with AI?
Start with the question buyers ask first: “Enterprise search vs a RAG chatbot: which should we build first?” NTSPL will answer it for your estate and give you a fixed-scope proposal.