Industry: Real estate, property operations, and enterprise services
Client: Large integrated real-estate group, identity withheld
Use case: Enterprise retrieval-augmented generation across real-estate documents and internal workflows
Delivered scope: Replacement of the major ingestion, parsing, retrieval, and model layers of an existing managed-cloud RAG system with domain-aware retrieval and an on-premise small language model
Disclosure: The client has been anonymized. The architecture and ₹20 lakh to ₹3 lakh monthly run-rate figures come from the Conscious Engines engagement record. The figures are presented as reported run-rates, not as an independently audited comparison of matched production periods.
1. Outcome at a Glance
The client's enterprise RAG system already worked. The problem was that its economics did not scale.
Conscious Engines rebuilt the major layers of the system around the client's real-estate language, document structures, and expected outputs. The replacement combined domain-specific ingestion and retrieval with an on-premise SLM, reducing dependence on usage-priced external model services.
The engagement record reports a reduction in monthly run-rate from approximately ₹20 lakh to approximately ₹3 lakh.
Key Outcomes
Approximately ₹20 lakh
Previous monthly run-rate
Reported engagement baseline; detailed cost composition is not public.
Approximately ₹3 lakh
New monthly run-rate
Reported rebuilt run-rate; full cost allocation is not public.
Approximately ₹17 lakh
Monthly reduction
Derived from the two reported run-rates.
Approximately 85%
Percentage reduction
Derived as (20 - 3) / 20.
Approximately 15%
New cost as a share of baseline
Derived from the reported run-rates.
| Measure | Reported result | Evidence boundary |
|---|---|---|
| Previous monthly run-rate | Approximately ₹20 lakh | Reported engagement baseline; detailed cost composition is not public |
| New monthly run-rate | Approximately ₹3 lakh | Reported rebuilt run-rate; full cost allocation is not public |
| Monthly reduction | Approximately ₹17 lakh | Derived from the two reported run-rates |
| Percentage reduction | Approximately 85% | Derived as (20 - 3) / 20 |
| New cost as a share of baseline | Approximately 15% | Derived from the reported run-rates |
| Annualized baseline | Approximately ₹2.40 crore | Mathematical annualization, not a claim of 12 months of realized spend |
| Annualized new run-rate | Approximately ₹36 lakh | Mathematical annualization, not a claim of 12 months of realized spend |
| Annualized difference | Approximately ₹2.04 crore | Scenario value if both monthly figures remain stable for 12 months |
| Deployment model | On-premise SLM with supporting enterprise infrastructure | Delivered architecture |
| Retrieval design | Domain-specific ingestion, parsing, and retrieval | Delivered architecture |
The reduction is economically significant. The previous run-rate was approximately 6.7 times the new reported run-rate. But cost alone does not establish success. A valid comparison must show that the rebuilt system processed a comparable workload and maintained an agreed level of answer quality, latency, availability, security, and human-review effort.
This distinction matters because RAG cost can be made artificially low by answering fewer questions, retrieving less evidence, using stale indexes, accepting lower quality, or moving work to employees. The credible claim is therefore not merely “85% cheaper.” It is:
The system's reported monthly run-rate fell by approximately 85% after the core RAG architecture was rebuilt. The public claim is limited to the reported run-rate change because matched traffic, quality, and full-cost data are not disclosed.
That wording is specific, strong, and defensible.
2. Why a Working RAG System Became Too Expensive
Enterprise RAG has several cost centers. The language-model API is visible, but it is not the whole bill.
total RAG cost = ingestion + parsing + embeddings + storage + retrieval + reranking + generation + orchestration + observability + human review
In a real-estate organization, documents are numerous, heterogeneous, and frequently versioned. A corpus can include agreements, technical reports, operating procedures, property records, vendor documents, project material, commercial correspondence, policies, presentations, and structured system exports. The exact corpus for this engagement is confidential, but the engineering problem is consistent across large property businesses.
Generic ingestion creates hidden waste
Real-estate documents do not behave like clean web pages. They contain repeated headers, schedules, annexures, scanned pages, tables, drawings, footnotes, cross-references, signature pages, and nested commercial definitions.
A generic parser can create several forms of waste:
- duplicate chunks from headers, footers, and repeated schedules;
- fragmented tables with lost row and column relationships;
- clauses separated from the definitions that control their meaning;
- page-level chunks that are too broad for precise retrieval;
- sentence-level chunks that lose document structure;
- repeated re-embedding when only a small part of a document changed;
- poor metadata that forces the retriever to search the entire corpus;
- longer retrieved contexts sent to expensive models.
Every weak ingestion decision creates downstream cost. Bad chunks increase the number of retrieval candidates. Poor retrieval increases reranking and context size. Oversized context increases generation cost and latency. Weak answers create more retries and human review.
A general model can be expensive in two ways
The first cost is direct: input tokens, output tokens, requests, provisioned throughput, and related managed services.
The second cost is architectural. If the model is expected to compensate for poor parsing, weak metadata, missing business rules, and generic retrieval, every request becomes larger and more complex than necessary.
For bounded enterprise tasks, the strongest general model may be used because the surrounding system is weak. That is an expensive substitute for better system design.
Cost should be normalized by useful work
Monthly cloud spend is a useful finance number, but it does not explain unit economics. A production system should measure:
- cost per query;
- cost per successfully completed retrieval;
- cost per grounded answer;
- cost per answer accepted without rework;
- cost per document ingested or refreshed;
- cost per active user;
- cost per business workflow completed.
The most useful denominator is usually cost per accepted answer or cost per completed business task. This prevents the team from declaring victory after reducing tokens while answer quality declines.
The problem in this case was therefore not simply that a cloud bill was high. The existing architecture coupled enterprise usage to a high variable cost without using enough domain structure to control retrieval and inference.
3. What Was Rebuilt
Conscious Engines replaced the expensive path through the system rather than applying a single optimization to the existing stack.
System at a Glance
Ingestion
Prepared real-estate content using domain-aware document handling.
Parsing
Preserved more useful document structure and metadata.
Chunking
Aligned retrieval units with the way the domain asks questions.
Retrieval
Ranked evidence using domain terminology and expected outputs.
Model
Shifted appropriate inference to an on-premise SLM.
Orchestration
Connected retrieval and generation as one controlled pipeline.
| Layer | Rebuilt function | Cost or quality mechanism |
|---|---|---|
| Ingestion | Prepared real-estate content using domain-aware document handling | Reduced noisy and duplicated content entering the index |
| Parsing | Preserved more useful document structure and metadata | Improved retrieval precision and reduced oversized context |
| Chunking | Aligned retrieval units with the way the domain asks questions | Reduced irrelevant passages and repeated model calls |
| Retrieval | Ranked evidence using domain terminology and expected outputs | Increased the chance that the first retrieval contained the right evidence |
| Model | Shifted appropriate inference to an on-premise SLM | Reduced recurring external token and API expenditure |
| Orchestration | Connected retrieval and generation as one controlled pipeline | Made routing, fallbacks, and monitoring possible |
| Deployment | Operated private inference inside the client's environment | Increased control over data movement, capacity, and runtime economics |
The high-level flow became:
enterprise documents and records
-> domain-aware parsing and metadata
-> structured chunks and embeddings
-> retrieval and optional reranking
-> on-premise SLM
-> grounded response for the authorized user
Domain-aware ingestion did more work before inference
The redesign moved intelligence earlier in the pipeline. Instead of sending loosely relevant content to a large model and asking it to resolve ambiguity, ingestion and retrieval prepared a smaller, better evidence set.
For real-estate content, a domain schema can include fields such as document type, property or project, counterparty, jurisdiction, department, effective date, version, confidentiality level, approval status, and source system. The exact schema used in this engagement is not disclosed. These are representative controls that make a production corpus easier to filter and audit.
The principle is simple:
better metadata -> smaller search space -> fewer retrieved tokens -> lower inference cost -> clearer citations
The SLM handled the bounded language task
The on-premise SLM did not need to reproduce the full capability of a frontier model. It needed to perform the client's recurring retrieval and response task at the required quality.
Suitable SLM work can include:
- classifying the user's question;
- selecting a retrieval route;
- identifying properties, dates, parties, document types, and other entities;
- reformulating a query using domain terminology;
- extracting structured fields;
- summarizing retrieved evidence;
- producing a response in the required format;
- abstaining or escalating when evidence is insufficient.
This is the practical reason a smaller model can create a large cost reduction. The system narrows the task before the model sees it.
On-premise did not mean “free”
Private inference replaces one cost structure with another. The complete post-migration cost should include:
- compute hardware or cloud-reserved capacity;
- depreciation or lease treatment;
- power, networking, and storage;
- inference-serving software;
- monitoring and security;
- model updates and evaluation;
- platform engineering and support;
- idle capacity and peak-capacity headroom;
- any external-model fallback traffic.
If the reported ₹3 lakh figure excludes material components that were included in the ₹20 lakh baseline, the comparison must be restated. A finance-approved total-cost model is more persuasive than a narrow API comparison.
The model was one part of the saving
It would be misleading to attribute the entire reduction to moving inference on-premise. The engagement scope describes a replacement across ingestion, parsing, retrieval, and model layers. Savings can come from fewer indexed duplicates, fewer retrieved tokens, fewer retries, smaller prompts, lower external-model usage, better caching, and more efficient model serving.
This is why enterprise AI cost reduction is an architectural problem, not merely a model-pricing exercise.
4. How Production Economics Should Be Measured
Before the headline becomes a public claim, the old and new systems should be compared over matched periods or through a controlled replay.
Normalize the workload
At minimum, record:
| Normalization variable | Why it matters |
|---|---|
| Queries per period | A quieter system naturally costs less |
| Active users | Adoption changes infrastructure and support load |
| Input and output length | Longer questions and answers change inference cost |
| Documents indexed | Corpus size affects parsing, embedding, storage, and retrieval |
| Documents refreshed | Update frequency affects recurring ingestion cost |
| Retrieval depth | More candidates and reranking increase compute |
| Availability target | Redundancy and peak headroom have real cost |
| External fallback rate | Some traffic may still use paid external models |
| Human-review minutes | Low automated cost can hide transferred labor |
A controlled replay is particularly useful. The team can run the same representative query set through both architectures, then compare cost and quality under the same traffic distribution.
Measure quality by failure mode
One blended “accuracy” number is inadequate for enterprise RAG. The evaluation should separate:
- retrieval recall: did the correct evidence enter the candidate set?
- ranking quality: did the system place the best evidence high enough?
- citation correctness: does each citation support the associated claim?
- groundedness: is the answer limited to retrieved evidence?
- completeness: did the answer include every required element?
- abstention: did the system refuse when evidence was absent or conflicting?
- access correctness: did retrieval respect the user's permissions?
- format compliance: did the response match the workflow's required structure?
- human acceptance: was the result usable without material correction?
The model should be evaluated on the client's own difficult questions, not generic benchmarks. A useful test set includes common queries, rare clauses, conflicting versions, scanned documents, tables, ambiguous entity names, missing evidence, and requests the user is not authorized to answer.
This is the same principle behind treating the enterprise evaluation set as a strategic asset.
Report a unit-economics scorecard
The engagement record supports the monthly run-rate figures. The operating measures required to normalize those figures are not public:
| KPI | Baseline | Rebuilt system | Change |
|---|---|---|---|
| Monthly total cost | ₹20 lakh | ₹3 lakh | 85% lower |
| Queries per month | Not publicly disclosed | Not publicly disclosed | Required to normalize cost |
| Cost per 1,000 queries | Not publicly disclosed | Not publicly disclosed | Useful infrastructure measure |
| Accepted-answer rate | Not publicly disclosed | Not publicly disclosed | Required quality measure |
| Cost per accepted answer | Not publicly disclosed | Not publicly disclosed | Primary economic KPI |
| p50 and p95 latency | Not publicly disclosed | Not publicly disclosed | Required service measure |
| Citation correctness | Not publicly disclosed | Not publicly disclosed | Required RAG quality measure |
| External-model fallback | Not publicly disclosed | Not publicly disclosed | Required routing measure |
| Human-review minutes | Not publicly disclosed | Not publicly disclosed | Required full-cost measure |
| Availability | Not publicly disclosed | Not publicly disclosed | Required service measure |
A fully normalized result would read: 85% lower monthly cost at comparable production traffic and accepted-answer quality. Until that matched comparison is available, “reported” and “approximately” are essential parts of the claim.
Extend the platform only after the base economics are verified
The rebuilt system can support additional real-estate and facilities workflows, but the following are roadmap opportunities, not claims about the delivered scope.
| Potential next workflow | Knowledge required | Decision or output | KPI |
|---|---|---|---|
| Lease and agreement retrieval | Approved agreements, amendments, definitions, dates | Cited answer and clause comparison | Review time, citation accuracy, missed obligations |
| Property operations copilot | SOPs, asset records, vendor contracts, work orders | Troubleshooting and next-action guidance | MTTR, first-time fix, escalation rate |
| Vendor and procurement intelligence | Contracts, invoices, service levels, performance history | Exception detection and supplier comparison | Leakage, invoice exceptions, SLA compliance |
| Project document assistant | Reports, approvals, drawings metadata, correspondence | Version-aware project answer | Search time, rework, decision delay |
| Energy and sustainability intelligence | Meter data, tariffs, equipment, occupancy, emissions factors | Driver analysis and recommended intervention | kWh per square foot, peak demand, CO2e |
| Voice-based field reporting | Engineer speech, site context, work order, evidence | Structured report and follow-up actions | Documentation time, completeness, closure time |
Each additional workflow should reuse the platform but own its evaluation set, permission rules, and business KPI. A common RAG layer does not make every use case equally safe.
5. What Enterprise Buyers Should Take Away
This case is useful because the original system was not a failed prototype. It was a functioning RAG implementation with an unsustainable cost structure. The improvement came from rebuilding the system around the actual task rather than continuing to pay a general-purpose architecture to absorb every inefficiency.
There are six practical lessons.
- Start with cost per accepted answer. Monthly spend is the symptom. Unit economics show whether the system becomes more efficient as adoption grows.
- Fix ingestion before buying more model capacity. Poor parsing and chunking increase every downstream cost.
- Use domain structure to reduce model work. Metadata, filters, document hierarchy, and expected-output schemas create cheaper and more reliable prompts.
- Give each model a defined job. A private SLM can handle frequent bounded work. Larger external models can remain available for justified exceptions.
- Compare full costs. On-premise compute, platform labor, fallbacks, and human review belong in the new baseline.
- Prove quality on the enterprise's own difficult cases. Generic benchmarks cannot establish production fitness for proprietary real-estate documents.
Production evidence standard
| Claim | Evidence required for a normalized comparison |
|---|---|
| ₹20 lakh previous monthly run-rate | Included services, invoices or cost ledger, tax treatment, and representative period |
| ₹3 lakh rebuilt monthly run-rate | Hardware and depreciation, power, hosting, support, engineering, and fallback inclusion |
| Approximately 85% reduction | Matched traffic and corpus, comparison dates, and finance approval |
| Comparable system quality | Retrieval, citation, groundedness, acceptance, latency, and availability results |
| On-premise SLM | Model and hardware description if publishable, capacity, utilization, and fallback policy |
| Business impact | User adoption, time saved, completed workflows, and avoided external work |
The immediate next step is not another feature. It is a matched production scorecard that turns an impressive cost number into a durable business case.
For a buyer evaluating the same problem, the useful diagnostic is straightforward: export 30 to 90 days of model, retrieval, ingestion, infrastructure, and human-review activity; build a cost-per-outcome baseline; replay a representative evaluation set; and identify which layers create cost without improving acceptance quality.
Related Conscious Engines research
- The intelligent property portfolio
- From work order to first-time fix
- The building that explains its energy use
- Cut the AI bill before it scales
- Build AI FinOps around cost per outcome
- Why one model is not an enterprise AI strategy
- Why your evaluation set is your AI moat
Sources and evidence boundaries
- Conscious Engines engagement record. It is the source for the previous architecture, rebuilt layers, on-premise SLM, and ₹20 lakh to ₹3 lakh monthly run-rate. The client's identity is intentionally withheld.
- The 85% reduction, ₹17 lakh monthly difference, and annualized figures are mathematical derivations from the internally reported monthly run-rates. They are not independently audited or evidence of 12 completed months at the new cost.
- Production traffic, answer quality, latency, availability, hardware cost, and human-review comparisons are not publicly disclosed. The headline should therefore be read as a reported run-rate change rather than a normalized independent audit.