The combination of LankaData and Chat2Find represents a fundamentally different approach to building artificial intelligence infrastructure for Sri Lanka. Rather than treating an AI chatbot as the starting point, the architecture begins with data. LankaData provides the structured data and knowledge foundation, while Chat2Find provides the language intelligence, reasoning, tool-use and conversational interface required to transform that information into usable answers and applications. Together, they form a data-to-intelligence architecture in which public information is collected, structured, indexed, retrieved, interpreted and ultimately delivered through artificial intelligence.
LankaData Net describes itself as an open data and intelligent access platform designed to transform fragmented Sri Lankan information into searchable and structured digital assets. Its repositories cover areas including legislation, judicial decisions, gazettes, parliamentary material, Central Bank statistics, economic indicators, pricing, taxation, business information and other public datasets. The platform states that its wider repositories comprise more than one billion datasets, while its LankaData Search layer provides contextual access to an archive of more than 90,000 Sri Lankan documents.
Chat2Find sits above this data infrastructure as the intelligence and interaction layer. LankaData’s own technical description states that its datasets are standardized and indexed and then wrapped with Retrieval-Augmented Generation models. A user query enters through Chat2Find, is interpreted and routed to the appropriate domain expert, relevant information is retrieved from LankaData repositories, and the resulting evidence is synthesized into a response.
The resulting architecture can therefore be understood as a national data pipeline that progresses through several logical stages: information acquisition, data plumbing, structured storage, knowledge representation, semantic retrieval, domain-specific RAG, language-model reasoning, application orchestration and user delivery.
From Public Information to Machine-Readable Data
The first layer is the information ecosystem itself. Sri Lanka’s authoritative information is distributed across ministries, departments, statutory bodies, regulators, courts, Parliament, the Central Bank, government publications, registries, financial institutions, public websites and other institutional repositories. Much of this information was historically designed for human consumption rather than machine consumption.
A national AI system cannot simply point an LLM at these websites and expect reliable answers. Government documents may exist as PDFs, HTML pages, scanned documents, tables, notices, gazettes, reports or structured datasets. Information may be duplicated across institutions, published at different times, amended later, or represented using inconsistent terminology.
LankaData’s role is consequently much closer to a data infrastructure provider than a conventional search engine. Its stated mission is to systematically compile, digitize and organize authoritative information across law, taxation, economics, business and public policy, converting fragmented information into clean, reliable and searchable digital assets.
This distinction is extremely important for AI. An LLM is fundamentally a statistical language system. It does not automatically know which version of a regulation is currently authoritative, which document superseded another document, whether a particular figure came from an official publication, or whether two different names refer to the same organization. Those problems belong to the data and retrieval architecture.
The Open PIN Concept
At the connectivity level, an architecture such as this requires an open information connectivity layer. The Open PIN concept can be understood as the interface between external information sources and the LankaData data platform.
Rather than having every downstream AI application independently connect to ministries, registries, databases and websites, the connectivity layer provides standardized mechanisms for acquiring information. These mechanisms can include APIs, document acquisition, web connectors, structured feeds, scheduled collection, incremental updates and change detection.
The importance of such a layer increases as the number of information sources grows. Without a common acquisition architecture, every application becomes responsible for its own collection process. LankaLAW would need one collection mechanism, LankaTAX another, LankaBIZ another and a news application another. That creates duplication, inconsistent versions and significant maintenance problems.
A shared connectivity layer allows information to be acquired once and reused across multiple downstream applications.
The architectural principle is therefore simple: the source should be connected to the data platform once, after which multiple services should be able to consume the resulting structured information.
Data Plumbing: The Invisible Infrastructure
The most important component of a national AI infrastructure may actually be the least visible to users: data plumbing.
Data plumbing describes the pipeline through which raw information becomes usable intelligence. Incoming material first needs to be ingested. It then needs to be stored, classified, parsed, normalized, enriched and indexed.
A document ingestion system may extract text from PDFs and HTML pages, identify document titles, dates, institutions and document types, and preserve the original source. Tables may need to be extracted into machine-readable structures. Scanned documents may require optical character recognition. Sinhala and Tamil documents introduce additional language-processing requirements because Unicode representation, typography and tokenization must be handled correctly.
The next stage is normalization. Dates need consistent representations. Institution names need to be standardized. Duplicate documents need to be identified. Different versions of the same legal instrument need to be associated. Entities such as companies, people, ministries, statutes and locations need to be resolved across datasets.
This is where entity resolution becomes important. If one source refers to an institution by its abbreviated name while another uses its formal name, an intelligent data layer should understand that the two records refer to the same entity. The same principle applies to company names, legislation, government agencies and geographical entities.
The output of this process is not simply a collection of documents. It is a structured information environment.
Structured Data, Documents and Knowledge Graphs
The resulting information can be represented through several complementary storage models.
Relational databases are useful for highly structured records. A company registry, for example, may contain company identifiers, names, registration information and associated attributes. Financial datasets can contain dates, indicators, values and classifications.
Document storage is better suited to reports, gazettes, judgments, circulars, press releases and other textual material where the original document structure is important.
A knowledge graph provides another dimension. Instead of simply storing text, it represents relationships between entities. A company can be associated with directors, subsidiaries, industries and regulatory records. A legal case can be associated with judges, statutes, cited cases and legal principles. A government publication can be associated with an institution, publication date, subject and subsequent amendment.
The knowledge graph therefore answers a different class of question from a conventional database. A database primarily asks, “What is stored in this record?” A knowledge graph can answer, “What is connected to this entity?”
Vector representations provide yet another layer. Documents and passages can be converted into embeddings that represent their semantic meaning. This allows a retrieval system to identify information that is conceptually related to a question even when the wording of the question differs from the wording of the source document.
The strongest architecture therefore does not depend on a single database. Structured databases, document repositories, knowledge graphs and vector indexes complement each other.
The Semantic Retrieval Layer
Once the information has been processed, the system needs to find the right information when a user asks a question.
This is the role of semantic retrieval.
Suppose a user asks a question about a tax provision without using the exact terminology found in the relevant legislation. A traditional keyword search may fail because the user’s words and the document’s words do not match exactly.
Semantic retrieval instead attempts to understand the meaning of the query and locate relevant passages, records or entities.
A sophisticated retrieval architecture can combine lexical search, vector search, metadata filtering, knowledge-graph traversal and structured database queries. These approaches can then be combined through a ranking mechanism that determines which pieces of evidence are most relevant.
This is the foundation upon which RAG operates.
Retrieval-Augmented Generation
Retrieval-Augmented Generation is the critical bridge between LankaData and the Chat2Find language models.
A conventional LLM generates an answer primarily from knowledge encoded in its parameters. That approach is powerful for general reasoning and language generation, but it has a fundamental weakness when answering questions about rapidly changing or highly specialized information.
RAG addresses this by retrieving relevant information at query time.
The process begins when a user submits a question. The system interprets the question, determines its subject and identifies the appropriate information domain. The retrieval engine then searches the relevant LankaData repositories. Relevant passages, records or structured information are returned to the AI layer. A ranking stage can further reduce the results to the most useful evidence. The system then constructs a context package containing the retrieved information and passes it to the language model.
The language model does not have to rely exclusively on what it learned during training. It can use the retrieved evidence to construct the response.
LankaData explicitly describes this architecture. It states that its curated datasets are wrapped with RAG layers and that queries are dynamically routed to domain-specific AI experts, which retrieve relevant information from LankaData repositories before the results are synthesized into a response.
This is one of the most significant aspects of the LankaData–Chat2Find architecture. The system is not simply attempting to make an LLM “know Sri Lanka.” It is creating an external knowledge infrastructure that an LLM can query.
Domain-Specific AI Experts
The next architectural layer is specialization.
Instead of building one enormous AI system that attempts to understand every subject equally, LankaData describes dedicated AI-powered domain experts for different datasets. The initial ecosystem includes areas such as law, taxation and business, with the architecture extending naturally into news, pricing, education and other domains.
A legal RAG system can be optimized around legislation, judgments, gazettes and legal terminology. A tax system can prioritize statutes, circulars, tax guidance and compliance information. A business system can operate over corporate, financial and market information.
This specialization improves retrieval because the system can constrain its search space and use domain-specific metadata.
It also improves prompt construction. The legal expert does not need the same instructions as the business intelligence expert.
The result is effectively a collection of specialized intelligence systems operating over a common national data infrastructure.
Chat2Find as the Intelligence and Orchestration Layer
Chat2Find provides the conversational interface and language intelligence that connects these specialized systems to users.
The architecture described by LankaData positions Chat2Find as the unified access point. A natural-language query enters the system, the language layer interprets the request and identifies the relevant domain, and the query is routed to the appropriate AI expert. Retrieved information is then synthesized into the final answer.
This makes Chat2Find more than a chatbot.
Technically, it can be understood as an orchestration layer. It can interpret intent, determine which domain should answer a question, call external tools, retrieve information, combine results and produce a natural-language response.
The publicly released Chat2Find-Instruct-v1 strengthens this architecture considerably. It is a 7-billion-parameter instruction-tuned model based on Chat2Find-CPT, itself based on Qwen3.5-7B according to the current Hugging Face model card. The model is designed for trilingual reasoning, tool calling and complex instruction following across Sinhala, Tamil and English.
The Chat2Find Model Stack
The model architecture itself has multiple layers.
At the foundation is Chat2Find-CPT, a continued-pretrained model designed to provide localized linguistic and cultural knowledge. Chat2Find describes this foundation as being trained on a large locally relevant trilingual corpus.
Above that is Chat2Find-Instruct-v1, which introduces instruction-following, reasoning and agentic capabilities.
The model’s training materials are particularly significant because they demonstrate an attempt to build localized AI infrastructure rather than simply wrapping a foreign model with a Sri Lankan prompt.
Chat2Find released a trilingual corpus containing approximately 255 million tokens across 279,248 conversations, covering Sinhala, Tamil and English and including code-mixed forms such as Singlish and Tanglish.
The subsequently released model documentation describes Chat2Find-Instruct-v1 as a 7B model with a 262,144-token native context window and a 279,260-record instruction dataset. The model supports trilingual reasoning and native tool calling.
The public Chat2Find technical description further states that the model uses QLoRA-based fine-tuning and 4-bit BF16 optimization, providing a path toward efficient inference and deployment.
This produces a model architecture in which localized data feeds continued pre-training, conversational data feeds instruction tuning, and the resulting model operates as an intelligent interface to external knowledge sources.
The Importance of Tool Calling
Tool calling changes the role of the LLM.
Without tools, the model produces text.
With tools, the model can become an orchestrator capable of interacting with external systems.
Chat2Find-Instruct-v1 is publicly described as supporting native agentic tool and function calling. The model can determine when an external API, database or search mechanism is required and then incorporate the returned information into its answer.
In a LankaData architecture, this capability can connect the language model to search APIs, structured databases, RAG retrieval services and other information systems.
This creates a closed intelligence loop.
The user asks a question. Chat2Find interprets it. The model determines what information is required. A tool retrieves that information from LankaData. The result returns to the model. The model reasons over the evidence and produces the response.
The LLM therefore becomes the control plane for an ecosystem of information services.
Trilingual Intelligence
One of the defining characteristics of the architecture is its focus on Sinhala, Tamil and English.
The challenge is not simply translating English into Sinhala or Tamil. Real Sri Lankan communication frequently involves code-switching, transliteration, local terminology, abbreviations and culturally specific references.
Chat2Find describes its model as being designed to handle Sinhala, Tamil and English together, including mixed-language patterns such as Singlish and Tanglish.
This is particularly important when the language model is connected to a national data infrastructure. A user should not have to translate a question into English before accessing Sri Lankan information.
The ideal architecture allows the question to remain in the user’s natural language while the retrieval system identifies the correct underlying data.
This means that multilingual capability needs to exist across the entire pipeline, not merely at the final response stage.
From Chatbot to AI Agent
The combination of RAG and tool calling allows Chat2Find to evolve from a chatbot into an AI agent platform.
A chatbot generally receives a question and generates an answer.
An agent can interpret a task, determine what information it requires, invoke tools, retrieve data, perform reasoning and generate a result.
For example, a complex business question could require information from several different datasets. The agent could identify the company, retrieve corporate information, obtain financial indicators, retrieve relevant market information and combine those results into a structured analysis.
The same architecture can support legal research, tax analysis, economic research, education and public-information discovery.
This is where the underlying LankaData infrastructure becomes particularly valuable. An agent is only as capable as the tools and information systems available to it.
Applications Above the Platform
The architecture enables multiple applications to operate over the same data and intelligence infrastructure.
LankaLAW represents the legal intelligence domain. Its architecture can connect legal documents, judgments, legislation, gazettes and other legal sources to retrieval and language models.
LankaTAX applies the same fundamental architecture to taxation.
LankaBIZ provides a business intelligence application over structured company, financial and commercial information.
Other Chat2Find ecosystem applications extend the concept into news, pricing and education. The key architectural advantage is that these applications do not each need to independently build a national-scale data acquisition system and language model.
They can share the underlying infrastructure.
This is the defining characteristic of a platform architecture.
The Data Flywheel
The combination of LankaData and Chat2Find can also create a data flywheel.
New public information enters the acquisition layer. It is processed and added to the repositories. Indexes are updated. RAG systems gain access to the new information. Users then interact with the AI applications.
In a mature system, interaction data can also help identify missing information, ambiguous queries, retrieval failures and new terminology.
This does not mean user conversations should automatically become training data. Strong privacy, consent, governance and data-quality controls are required. But from a platform perspective, usage can identify where the underlying knowledge infrastructure needs improvement.
Over time, the system can therefore become better at both information retrieval and language interaction.
Provenance and Trust
For a national information platform, accuracy is not sufficient. Traceability is equally important.
An AI-generated answer should ideally be connected to the underlying evidence. This is particularly important for legal, taxation, financial and government information.
LankaData emphasizes authoritative sourcing, validation, traceability and institutional reliability. Its stated data sources include official publications, statutory instruments, policy documents, judicial records, economic reports and regulatory disclosures.
A technically mature implementation should therefore maintain provenance metadata throughout the pipeline.
The system should know where a record originated, when it was collected, what transformations were applied, which version is current, which document it belongs to and which retrieval result was used to construct an answer.
This creates a chain of trust:
Source → Ingestion → Transformation → Repository → Retrieval → RAG Context → Generated Answer.
The closer the architecture gets to preserving this complete chain, the more suitable it becomes for professional and institutional use.
Data Governance and Security
The technical architecture cannot be separated from governance.
A platform aggregating national legal, business, financial and public information needs strong access control, authentication, authorization, audit logging, data classification, retention policies and security monitoring.
Different datasets may have different licensing and access conditions. Some information may be public, while other information may require controlled access.
The architecture therefore needs to distinguish between the data acquisition layer, the internal data layer and the external application layer.
API gateways, authentication systems and role-based access controls can protect the interfaces between these layers.
Encryption should protect sensitive information in transit and at rest. Audit systems should record important access and administrative activity. Backup and disaster-recovery mechanisms are necessary because the platform increasingly becomes infrastructure rather than simply a website.
Infrastructure and Deployment
Behind the logical architecture is the physical computing infrastructure.
The data-processing layer requires storage and compute capacity for document processing, transformation, indexing and database operations. The AI layer requires GPU infrastructure for model inference and potentially training or fine-tuning. Vector search and knowledge-graph workloads require their own specialized storage and indexing mechanisms.
Containerization can separate services and allow individual components to scale independently. An API gateway can provide a common entry point for applications and developers. Load balancing distributes traffic across service instances.
Caching can reduce repeated retrieval operations and improve latency for frequently requested information.
Observability becomes increasingly important as the architecture grows. Metrics, logs and traces should make it possible to identify whether a delay occurred in the API layer, retrieval engine, database, RAG pipeline or model inference layer.
The result is a distributed AI infrastructure rather than a single application server.
The Architecture as a National Information Operating System
The most important way to understand LankaData and Chat2Find is not as two websites.
They can instead be viewed as two major layers of a broader national information operating system.
LankaData provides the information substrate.
Data plumbing transforms raw information into machine-readable knowledge.
Search, databases, knowledge graphs and vector indexes make that knowledge retrievable.
RAG converts retrieval into grounded context.
Chat2Find provides language intelligence, reasoning and agentic orchestration.
Applications such as LankaLAW, LankaTAX and LankaBIZ turn that intelligence into domain-specific services.
The user ultimately sees a simple conversational interface, but underneath it is a multi-layer information architecture.
The complete conceptual flow is therefore:
Sri Lankan information → Open connectivity → Data ingestion → Data processing → Structured repositories → Knowledge graph and vector indexes → Search and retrieval → Domain RAG → Chat2Find LLM → Tool and agent orchestration → AI applications → Users.
Why the Architecture Matters
The strategic significance of this architecture is that it separates knowledge from intelligence.
The knowledge layer can continue to grow and change without requiring the language model to be completely retrained every time a new government report, regulation, judgment or economic statistic becomes available.
Conversely, the language model can improve without requiring the entire data infrastructure to be rebuilt.
This separation is one of the central advantages of RAG-based AI architecture.
A model provides reasoning and language capability. The external data infrastructure provides current and authoritative knowledge.
LankaData and Chat2Find therefore create a complementary architecture in which the data platform and AI platform strengthen each other.
The April 2026 announcement by LankaData explicitly described the collaboration as combining LankaData’s structured repositories with Chat2Find’s language-model capabilities and RAG technology to create an intelligent AI layer for accessing public data.
The Long-Term Architecture
The logical end state is considerably larger than a search engine or chatbot.
The same infrastructure can become a platform through which government information, legal knowledge, economic statistics, corporate data, education resources, market information and other national datasets become programmatically accessible.
Developers can consume APIs.
Researchers can query structured datasets.
Professionals can use domain-specific AI experts.
Businesses can integrate intelligence into their own applications.
Citizens can ask questions using natural language.
Government institutions can use the system for information discovery and decision support.
AI agents can use the same infrastructure programmatically through tools.
The significance of this model is therefore not merely the creation of a Sri Lankan chatbot. It is the construction of a reusable data and intelligence infrastructure in which authoritative national information becomes accessible to both humans and machines.
LankaData and Chat2Find represent two halves of the same technical proposition.
LankaData addresses the problem of fragmented information. It collects, digitizes, structures, indexes and maintains Sri Lankan knowledge resources across multiple domains.
Chat2Find addresses the problem of interacting with that knowledge. Its localized language models provide trilingual understanding, reasoning, instruction following and tool-use capabilities, while its RAG architecture allows the model to access external knowledge rather than relying exclusively on information encoded in model parameters.
The resulting architecture is a data-to-intelligence pipeline rather than a conventional AI application.
At the bottom is Sri Lanka’s information ecosystem. In the middle are the data plumbing, repositories, search systems, knowledge graphs, vector indexes and RAG experts. Above them sits Chat2Find’s model and agent layer. At the top are domain applications and user-facing services.
That architecture creates a potentially powerful national platform because every new dataset can increase the usefulness of the intelligence layer, while every improvement in the intelligence layer can increase the value of the underlying datasets.
The fundamental proposition is consequently simple:
LankaData provides the knowledge. RAG provides grounded retrieval. Chat2Find provides intelligence. AI applications provide utility.
Together, they form a foundation for transforming Sri Lanka’s fragmented public information ecosystem into an interconnected, searchable, machine-readable and AI-accessible national knowledge infrastructure.
