Lab

Semantic search: what embeddings can retrieve

Short answer

Semantic search connects texts by meaning, even when they use different words. It can recover context expressed in another way. It cannot establish that a note answers the question, that a decision remains valid or that the reader is authorised to access it.

Different wordings lead to a shared note whose relevance still needs checking.

From wording to similarity

An embedding is a numerical representation of text. Comparing representations helps identify documents related to a question. In a fictional example, “Why was this equipment retained?” may connect with a note explaining a retention decision without sharing the exact wording.

Keyword search remains useful for exact references, names and specific terms. Semantic search addresses another difficulty: people asking a question may use different vocabulary from the source. Combining the approaches requires measuring each contribution rather than assuming that a more elaborate system necessarily performs better.

A semantic neighbour is not an answer

Two notes may discuss the same supplier but concern different assets. An old decision may be highly similar to the question despite having been superseded. A rejected proposal may contain precisely the amount being searched for.

Similarity therefore measures neither truth, authority nor temporal validity. A high-ranking note remains a candidate to read. Moving from candidate to answer requires its provenance and qualifications to survive retrieval.

The Lab’s alert experiments exposed this limit too. A decisive fact buried in a meeting record could fail to appear while a neighbouring note attracted attention. The diagnosis must examine document granularity and coverage, rather than simply tune a similarity threshold.

Multilingual ability needs product-level evaluation: scope and limits

The BGE-M3 model card describes multilingual capabilities. That model property does not establish equal retrieval quality for every language in Nawame. Corpus composition, text length, business terminology and question wording all affect the outcome.

A useful evaluation separates searches within one language from searches where the question and source use different languages. It also tests unanswered questions and false matches. Connecting languages is a capability to assess in context, rather than a guarantee of translation quality or fidelity.

Longer questions are not always clearer

Putting every detail into a query can dilute its main subject. A Lab campaign encountered this while investigating links between sources. A search that depended on an earlier result benefited from identifying the intermediate fact instead of asking everything in a single sentence.

Document chunking raises a similar issue. A representation spanning several topics may obscure a short decision. Smaller units can help, but isolating a sentence can also remove its condition. This trade-off motivated the Lab’s passage experiments.

The consequence: evaluate the full chain

For Nawame, the embedding engine is a shared component for retrieving context. It replaces neither access rights, version history nor the authorised assistant’s interpretation. Its effectiveness must be judged alongside relevance, abstention, response time and processing cost.

Multi-step retrieval and effective dates show why useful memory needs more than a ranking of similar documents. Meaning helps locate material; evidence establishes what can reasonably be concluded. The distinction lets us improve retrieval without presenting proximity as certainty.

Related reading: When semantic search finds a precedent that never existed and AI memory vs RAG: they are not the same thing.

Your next prompt should not have to contain your whole story.

Start building a memory your AI tools can find again when it matters.

Start free