1. Home
  2. AI development
  3. Semantic search

Semantic search

Semantic search that finds the right document, even in different words

Semantic search matches what a person means with what a document says, so a search for "refund after 30 days" finds the policy titled "Return window". We build it over your internal documents, past tickets, contracts or a parts catalog, with exact-code matching alongside and permissions respected.

Get the free audit Three ways to search
  • Prices published on the site
  • First working version in two weeks
  • Support month to month, cancel any time

Search by meaning, over the files you already have

Keyword search finds the words you typed. Semantic search converts each passage of text into an embedding, a list of numbers that places similar meanings close together, and compares your question with those. The embeddings live in PostgreSQL with the pgvector extension or in a vector database, next to the original files.

Semantic search returns the right document. Private AI over your documents goes a step further and writes an answer with citations. Many projects start with search alone: a ranked list is cheaper, easier to check, and the answers can be added later on the same index.

Keyword, semantic or hybrid

Keyword search

Matches the words typed

Finds documents that contain the exact words or codes. Fast and predictable, and it misses a paraphrase.

  • Finds paraphrases and synonyms: Low
  • Finds exact codes and part numbers: High
  • Setup effort: Low

Best for: Part numbers, invoice numbers, section numbers

Semantic search

Matches the meaning

Finds the right passage even when it uses other words. It can miss an exact code or rank a near match above it.

  • Finds paraphrases and synonyms: High
  • Finds exact codes and part numbers: Low
  • Setup effort: Medium

Best for: Long text that people ask about in their own words

Hybrid with reranking

Both, then a second sort

Runs keyword and semantic search together, then a reranking model sorts the combined results by how well each one answers the question.

  • Finds paraphrases and synonyms: High
  • Finds exact codes and part numbers: High
  • Setup effort: High

Best for: Catalogs and ticket history where people use codes and plain words

We start with the simplest option that passes your test questions and add the next one only when the test shows it is needed.

Sources we search, and what we add around them

Internal documents

Policies, procedures and handbooks, found by the question an employee would ask, whatever the file happens to be called.

Past support tickets

Finds how a similar problem was solved before, even when the customer described it differently.

Contracts

Finds the clause on termination or payment terms across hundreds of agreements, and shows the page it came from.

A parts or product catalog

Staff find an item by describing it or by typing its code, from the same search box.

Permissions respected

Results are checked against the access rules of their source. We write down how each source's permissions carry over and test with accounts at each access level.

Kept current

Changed files are re-indexed on a schedule we agree, so search returns the current version.

When semantic search pays off, and when it does not

A good fit when

  • You have thousands of documents or years of tickets, and keyword search keeps missing
  • Staff ask in their own words and the documents use other terms
  • Contracts or manuals are long, and the answer sits in one paragraph

Something simpler will do when

  • You have a few hundred documents with good folder and file names: the search built into Google Drive, SharePoint or your file server is enough
  • You look up exact numbers that your system already finds, such as an order number in your accounting software
  • You need a written answer instead of a list of documents, which is private AI over your files

Priced by the number of sources

One process: $2,500 to $6,000, 2 to 4 weeks

One workflow running end to end, or one internal tool.

Connected systems: $6,000 to $15,000, 4 to 8 weeks

Several systems made to work as one: integrations, roles, reports.

Search over a single source, such as one document library or one ticket system, usually lands in one process. Search across several sources, with permissions carried over from each, usually lands in connected systems. Indexing the documents has a one-time model cost, and each update and search adds a smaller one; we estimate both in the written review.

Care $300 to $500 a month. Active $750 to $1,500 a month. AI model usage, GPU time and hardware are paid to the provider, on accounts in your name.

Get the free audit See all prices

Questions about semantic search

How is semantic search different from the search we already have?

Most file search matches the words you type, though some tools now add natural-language search, so check yours first. Semantic search also finds passages that mean the same thing in other words, and we pair it with keyword matching so exact codes still work. The written review says whether your documents would benefit.

Will people see documents they should not?

Not if the permissions are carried over, which is part of the scope. Each result is checked against the access rules of the source system before it is shown. Where a source cannot express per-document rules, we split the index by group, for example HR files apart from the rest, and test with accounts that have different access.

Where are the embeddings stored, and do our documents leave our server?

The embeddings go in PostgreSQL with pgvector, or in a vector database, on your own server or cloud account. Creating them takes an embedding model. If that is an outside service, the document text goes to that provider, and we tell you which one. An open-weight embedding model can run on your own server when nothing may leave.

How do you know it finds the right document?

We write 20 to 50 real questions with you, each with the document that should come back, and count how often it lands in the top five results. We rerun the same set after every change, so an improvement for one kind of question does not quietly break another.

How many documents can it handle?

Hundreds of thousands of passages fit on one ordinary PostgreSQL server with pgvector. Much larger collections or heavy traffic may call for more hardware or a dedicated vector database, and we say which in the audit.

Do our files need cleaning first?

Less than you expect. Scanned PDFs need their text read first, which multimodal AI covers. Duplicates and old versions are filtered or ranked lower, and you point us to the folders and systems to index. Anything that has to be cleaned first goes into the written scope.

Related services

Tell us what's slowing you down.

You get a free written review by email within two business days.

Only your email and the task are required. No passwords or client data needed.

No mobile information will be shared with third parties or affiliates for marketing or promotional purposes.