Answers from your product’s knowledge.

Make your knowledge usable inside the product

A retrieval-augmented system finds relevant material before a model drafts an answer. It can help staff or customers navigate documents, policies and product information. Citations support review; they do not guarantee that an answer is correct.

The first decision is which questions the system is responsible for and which documents may answer them. A corpus assembled by exporting everything available will also retrieve superseded material that reads as current; one limited to the documents that are current and authoritative is easier to evaluate.

What the implementation can include

A pipeline for agreed document formats, indexing and retrieval, permission filtering, source links and refresh behavior. Hybrid search and reranking are options when your corpus and evaluation justify them.

Ingesting the documents you already have

A corpus usually mixes exported help-center HTML, PDFs from word processors or print layouts, spreadsheets, wiki pages and email threads. Each format needs a parser that preserves reading order, headings and tables, and discards navigation, repeated headers and boilerplate that would otherwise be retrieved.

Parsing also produces the metadata retrieval needs later: title, section path, effective date, source system and the permissions attached at source. Reconstructing those from chunk text afterwards is unreliable.

How documents are split for retrieval

Splitting determines what can be found at all. A passage that answers a question is only retrievable if it survives as a coherent unit: fixed-size splitting cuts tables, clause numbering and multi-paragraph answers apart, so the half carrying the answer loses the half that gives it context.

Structure-aware splitting follows the document instead. Headings become boundaries and travel with each passage as a section path, so a chunk keeps the context its page supplied. Short sections are merged with their neighbors and long ones are divided at a paragraph boundary.

Tables and scanned pages

Pricing, specification and policy documents often hold their answers in tables, and generic parsing handles tables poorly: flattened into a sequence of cell values, a table loses the row and column relationships. A table can instead be kept whole as one passage, with its headers repeated per row. Scanned pages need optical character recognition, whose quality varies with the scan, so we test it on a sample.

Embeddings and where the index lives

Embedding choice affects retrieval quality, cost and how much work a later change requires. A general-purpose model may not separate terms that are distinct in your field, which is visible only in evaluation against your own corpus rather than in published provider benchmarks. Changing the model later means re-embedding everything.

An existing Postgres instance with a vector extension is often sufficient; a dedicated vector database is warranted when corpus size, filtering or query volume require one.

Hybrid retrieval and reranking

Dense vector search finds passages whose embedding is close to that of the question, approximating topical relatedness rather than word overlap. It performs poorly where an exact string is required: part numbers, error codes and field names. Keyword search handles those and misses paraphrase, so retrieval runs both and merges the results through rank fusion, with filters for type, date validity and access applied first.

A reranking step then scores each candidate against the question directly. That is usually more accurate than comparing precomputed vectors, and it costs more, so it runs over tens of candidates rather than the whole corpus. The number of passages passed to the model trades recall against prompt cost and against reliability on long prompts, since passages in the middle of a long prompt are used less than those at either end on many models.

Citations and missing evidence

A resolving citation confirms that the passage was retrieved, not that it supports the statement. Whether it does is what the evaluation set scores.

Where the best retrieved passages fall below the agreed support threshold, the system states that the material is not available. It offers related documents or a route to a person rather than answering from general model knowledge. Similarity scores are not comparable across queries, so the threshold is set on reranker scores or an explicit support check, calibrated against the evaluation set.

Retrieval that respects permissions

Access control is therefore applied as a retrieval filter rather than after scoring. Filtering after retrieval returns a shorter list from a search that already read restricted content, and instructing the model not to use a passage it has been given is not an access control.

A permission captured at ingestion goes stale when the source changes, so permission changes are re-synchronized on the same schedule or change feed as content, and the lag between the two is agreed rather than left implicit.

Freshness, versions and re-indexing

The pipeline also needs content hashing so unchanged documents are not re-embedded, deletion handling so withdrawn documents leave the index, and a record of when each document was last indexed.

Evaluate the questions your users ask

Use representative questions, missing-answer cases, conflicting documents and access boundaries. Agree the evaluation set and acceptance criteria before implementation. Freshness, model quality and source quality all affect results.

The set is built from questions people actually ask, taken from support tickets and search logs, and it has to include questions the corpus does not answer, which answerable questions alone will not surface. Each entry records the question and the passages that should have been retrieved, so retrieval and generation are scored separately. Scoring runs on every change, because a change at one stage can regress another.

What a retrieval project costs

Every engagement starts with a fixed-price discovery sprint at USD 2,500 over two weeks, which produces the architecture, initial screen designs, acceptance criteria, a running-cost model and a fixed price for the build. The sprint fee is deducted from the build.

Indexing is charged per document version, and it recurs whenever documents change or the corpus is re-indexed against a different embedding model.

Deciding whether a requirement needs retrieval at all: prompting, retrieval or fine-tuning → · what it costs to run →

Part of the AI Feature Implementation Sprint → · Explore our work → · Read why naive RAG fails →