A language model can produce fluent Indonesian sentences without having adequate evidence about Indonesia. The gap is most apparent in questions whose answers depend on time, location, or precise wording: public service requirements, regulation numbers, recently changed prices, and events in a particular region.
For questions of this kind, the problem is not simply one of adding knowledge to the model. The system must find relevant documents when a question is asked, select evidence worth bringing into context, then show where it came from. Ison Search handles this sequence for IsonAI. This note covers the reasoning behind each stage, the consequences of those choices, and the failures that remain possible.
Why retrieval is needed
Knowledge stored in model weights is static. It can help explain general concepts, but offers no assurance that a rule remains in force, a schedule has not changed, or a service is available in a particular regency. Fluency can also conceal this gap: an answer may sound complete despite being built from outdated recollections or estimates.
Retrieval, or the search for evidence, moves part of the problem to the point when the question is asked. The system searches for documents that might support an answer; this is not a matter of copying a ready-made response. The distinction matters. Evidence may be spread across several pages, answer only part of the question, or conflict. Even after it has been found, it still has to be read and assembled.
Retrieval is therefore not a shortcut to the truth. It is a mechanism for reducing reliance on the model's memory and making the factual basis of an answer visible. The final result remains limited by the coverage of the index, the query formulation, ranking, and the model's ability to use context without going beyond what the sources state.
What Ison Search searches
The Indonesian web has no single type of source that is always the most appropriate. Government sites matter for the text of regulations, official requirements, and fees. National and regional media record changes and implementation on the ground. Institutional sites provide information from organisations working in the relevant field, while academic publications and university repositories are useful when a question requires methods, data, or a more extensive analysis.
These differences make authority contextual. For the wording of a provision, the official document should usually come first. To determine whether it has only recently been announced or how it is applied locally, a newer source or one closer to the event may provide evidence absent from the official page. An official source is not automatically current; the newest source is not automatically correct either.
Ison Search therefore looks for a match between the question, the document's content, the source it came from, and its publication date. The aim is not to create a list of publishers trusted without qualification, but to present a body of evidence suited to the claim being answered. A source may be highly authoritative for one claim and irrelevant to another claim in the same question.
Why hybrid search
Before search begins, the query is normalised and interpreted so that spelling variants, abbreviations, and its important elements can be preserved. This representation is used to create an embedding, a vector that captures semantic proximity. The query then enters two parallel lanes: lexical search and vector search.
BM25 is the lexical lane. It uses an inverted index to find documents containing query terms and weighs how informative those terms are. This approach works well for institution names, technical terms, regulation numbers, and character sequences that must match exactly. Its weakness appears when the user and publisher express the same thing in different words: the right document may be missed because its paraphrase shares too few terms with the query.
Vector search compares the query embedding with document embeddings. Its nearest neighbours are found with KNN through an ANN index, such as HNSW, keeping search practical across large collections. This lane is better at capturing paraphrases and semantic relationships that are not apparent from shared words. Semantic proximity, however, is not evidence of factual accuracy. A document may cover a very similar topic but come from another jurisdiction, be out of date, or lack authority over the claim being sought.
Both lanes are retained because they fail differently. BM25 preserves term precision; vector search improves recall across language variations. Running only one would mean accepting one kind of blindness as a permanent property of the system.
Ison Search.hybrid retrieval for Indonesian sources- BM25 for term weighting
- Inverted index (FTS)
- Tokenisation & Indonesian word variants
- Query & document embeddings
- KNN with ANN (HNSW)
- Cosine similarity, retrieve top-k
- Authority: official/primary source
- Freshness: publication date
- Top-N context provides the basis
- Claims follow the sources
- Citations are attached to the answer
From candidates to context
The outputs of the two lanes cannot be combined directly. BM25 scores and vector similarity measures come from different scales and assumptions; adding them would assign meaning to numbers that are not actually comparable. Ison Search uses RRF (reciprocal rank fusion) to combine the lists by ranking position. Documents that rank highly in both lanes receive stronger support, while documents found by only one lane can still remain in contention.
The combined list may still contain duplicate pages or numerous documents from the same domain. Dedup removes repeated results, then a per-domain limit prevents one publisher from dominating. Both are controls on the candidate set, not ranking signals or judgements that a domain is of poor quality. This choice deliberately trades a degree of recall—some useful documents may also be removed—for greater diversity of evidence and room for other sources.
Once the candidate space has narrowed, query and document pairs are reassessed by a cross-encoder. Unlike vector search, which compares separate representations, a cross-encoder reads both together and can assess their relationship in greater detail: whether a document genuinely answers the requested requirements or merely mentions the same topic. This assessment costs more, so running it across the entire index would make little sense. The initial search stage is responsible for coverage; reranking provides precision over an already limited set of candidates.
During reranking, content relevance is considered alongside authority and freshness. Their weighting depends on the question. A primary document should take precedence for the wording of a regulation, but an old date may indicate that the procedure has changed. A recent report helps with current events, but recency is no substitute for verification. After these considerations, only the highest-ranked documents—the top-N—are brought into context. This limit keeps the context focused, at the risk of excluding supporting evidence beyond the cut-off.
From context to answer — and its limits
IsonAI then constructs an answer on the basis of that top-N context, a pattern known as RAG (retrieval-augmented generation). At this stage, documents are not simply summarised one by one. Information spread across them must be connected, differences in scope preserved, and claims kept within the available evidence. The sources used are linked as citations.
Citations make an answer verifiable, not automatically correct. The model can still misinterpret a sentence, combine two provisions that apply under different circumstances, or draw a broader conclusion than its sources support. Links allow readers to assess these errors; they do not remove them.
Failures can also occur before the answer is constructed. The right document may not yet be in the index, the query may emphasise the wrong terms, ANN may miss a neighbour that should have been found, or the candidate limit may exclude evidence that proves important. The context that remains may be incomplete, outdated, or contradictory. In such cases, the answer should retain the uncertainty and identify the conflict instead of flattening it into a single certainty.
A good output therefore depends on more than the ability to compose language. It depends on an auditable chain of decisions: what evidence was found, why certain evidence was prioritised, what was not retrieved, and which claims are genuinely supported by the sources. This chain does not guarantee that every answer is correct, but it makes the source of an error more likely to be found and tested.
IsonAI is available through isonai.net and the IsonAI app for iOS.