# Search, rebuilt: how Falconer finds the right answer

> Company search is hard: the corpus is too large to browse and too small to learn ranking from. Here's how Falconer's search page and answer engine find the right answer.

- Date: 2026-06-23
- Author: Aryaman Agrawal
- Tags: engineering, search, retrieval, agents

---
![Falconer search: many connected sources converging into one query and one ranked answer](/images/notes/search-rebuilt-hero.png)

Bad search is one of the most frustrating user experiences. You know something exists, but you can't find it. You sift through old versions, poor rankings, or irrelevant results.

Company search is hard because the corpus is too large to browse and too small to build ranking signals from. Google's search algorithm is great because it has billions of search results per day to learn from.

The most popular document in an enterprise may only have 50 views that week. There's a long tail of important, but one-off searches.

## Search across everything

The key piece to great company search is structured content. Falconer's search is unique because it sits on top of an intelligent knowledge graph and structured document taxonomy.

The beauty of the knowledge graph means [industry-leading results in answer quality](https://falconer.com/benchmarks).

You have two UIs to search within Falconer:

- **A search page** for finding hits across every data source. Think traditional Google search over your company data.

- **An answer engine** that generates a natural language response instead of a list of links. Think ChatGPT over your company data.

Search across your native Falconer docs and every source you connected: Notion, Confluence, Google Drive, Linear, GitHub, Zendesk, and more.

![Falconer search results page: a query for "auth" with type, last-edited, author, and source filters, showing ranked results with highlighted snippets and the path to each document](/images/notes/search-rebuilt-srp.png)

## From results to answers

Instead of a ranked list of results, it returns one synthesized, cited answer drawn from your own content.

- **Plain-language questions work.** Keyword and vector search run together, so "how do we handle auth?" finds the right document even when it shares none of your exact words.

- **Code and docs together.** A single question can pull from the implementation and the design note beside it, not just one or the other.

- **Summaries with citations.** You get an answer with sources attached, instead of six documents to reconcile yourself.

- **Grounds your own agents.** Point an MCP client or the API at Falconer and your tools answer from your high-quality content.

![Three properties of the answer layer: docs and code together, semantic matching from a fuzzy query, and one cited answer instead of a reading list](/images/notes/search-rebuilt-cards.png)

## How it works

Two retrieval shapes, one index.

**The search page** runs full-text search on OpenSearch:

- Matches across titles, body text, and identifier-aware fields, with fuzzy matching for typos and a floor that requires most of your terms to land.

- Ranks by relevance, and you can sort by relevance or by what changed most recently.

- Permission-checks every result at query time and again before it reaches you, so search never shows you a document you can't open.

**The answer layer** retrieves the same documents with a hybrid query:

- Keyword (BM25) and vector (semantic) run together and fuse, then results are weighted by **document authority**, so a source-of-truth doc outranks a private scratch note.

- Short documents expand to their full text, and neighboring passages come along so the answer has room to reason.

- If the embedding step fails, the query falls back to keyword search rather than returning nothing.

![Two layers over one index: the search page runs full-text BM25 ranked by relevance into ranked results; the agent and API run hybrid keyword plus vector retrieval weighted by document authority into a cited answer](/images/notes/search-rebuilt-flow.png)

## How it compares to Glean

Glean is the closest comparison: enterprise search that connects to your tools and answers across them. The difference is what each does with what it finds.

| Capability            | Falconer                                                              | Glean                                  |
| --------------------- | --------------------------------------------------------------------- | -------------------------------------- |
| Time to value         | Connect sources and search in minutes                                 | Contract plus a multi-week rollout     |
| Keeps content current | Drafts doc updates as code ships, keeping the source of truth current | Surfaces what is there, fresh or stale |

The gap shows up before you run a single query. Glean is a contract and a multi-week rollout; Falconer connects and searches in minutes. And once you are in, Falconer doesn't just rank what already exists. As code ships, the knowledge layer drafts updates to the docs it touches, so your source of truth stays true instead of quietly going stale.

## What you get

- **Less time hunting.** One search, every source, the right result near the top.

- **Results you can trust.** The answer layer weights for document authority, surfacing the current source of truth, not whatever matched the most keywords.

- **Answers when you want them.** Ask in plain language and get a cited answer drawn from your own content.

- **Lighter agents.** Because the first retrieval is right, agents reading from Falconer finish in a turn or two instead of widening the net and pulling more context.