You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 09ff227
Browse filesBrowse the repository at this point in the historyBrowse files
Copy file name to clipboardExpand all lines: website/docs/reliability-context-quality.md
+1-5Lines changed: 1 addition & 5 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -73,10 +73,6 @@ threshold stays dark.
73
73
74
74
The index is built once and paid for once. Every query after that embeds only the question — a few tokens against a corpus that is already vectorized — so asking is nearly free. It is the same bargain as an index on a database table: pay once to organize the data, then read a short list instead of scanning the table. That is the economics that make retrieval worth the setup: a cheap retriever hands the LLM reader a handful of passages, and the reader spends its attention on those instead of the whole corpus.[^dpr] Measured against feeding everything into a long context, retrieval is the cheaper, more attention-efficient choice — not automatically the more correct one.[^rag-efficiency]
75
75
76
-
:::warning The Agentic Coding Cost Trap
77
-
Most providers charge a per-token surcharge past a context threshold — often double the input rate, well inside the advertised window. For one-shot retrieval tasks that premium buys tokens in the model's strong accuracy zone; for agentic coding it buys tokens in the forgotten middle, where attention is weakest and requirements go to die. Retrieval keeps you under the threshold; feeding everything in lives above it.
78
-
:::
79
-
80
76
## Deep Research: Search, Learn, Iterate, Cite
81
77
82
78
A single semantic search answers a narrow question against a corpus you already indexed. Deep research answers a question whose answer is spread across sources nobody has read yet. Strip away the branding and every implementation is the same two phases: **exploration** gathers the candidate pieces of reality, and **synthesis** turns them into an answer.
@@ -96,7 +92,7 @@ for precision — narrow, ordered, answer-shaped.
96
92
97
93
Most chat interfaces already ship it: flip on deep research, ask the sprawling question, and wait. What comes back is written, cited, and compact. Underneath, it is the [grounding phase](./workflow-grounding.md#phase-1-grounding) of the [four-phase workflow](./high-level-methodology.md) — a sub-agent that explores, filters, and returns a distilled answer instead of the pile it read — and it is a [workflow agent](./workflow-agents.mdx): deterministic code owns the loop, the budget, and the stop condition, while the model is called at bounded junctions to judge what is missing and to write. The orchestrator receives the conclusion and asks the next question; the raw pile never reaches it.
98
94
99
-
Those products aim it at the web, where the retrieval tools already exist. The technique is the part that travels: the same two phases run over a codebase, a ticket archive, or a shelf of documents, reading symbols and line ranges, or sections and links, instead of search results. The reading changes; the loop does not. Length bends the same way — synthesis decides how much of what it gathered earns the page, so the same run returns a paragraph or a long report. [ChunkHound](https://chunkhound.github.io)[disclosure]— a sister project of this book — runs that same loop over code, git history, and the web, locally on the developer's laptop, so the corpus never leaves the machine.
95
+
Those products aim it at the web, where the retrieval tools already exist. The technique is the part that travels: the same two phases run over a codebase, a ticket archive, or a shelf of documents, reading symbols and line ranges, or sections and links, instead of search results. The reading changes; the loop does not. Length bends the same way — synthesis decides how much of what it gathered earns the page, so the same run returns a paragraph or a long report. [ChunkHound](https://chunkhound.github.io) — a sister project of this book — runs that same loop over code, git history, and the web, locally on the developer's laptop, so the corpus never leaves the machine.
100
96
101
97
A run is a different order of cost: minutes rather than milliseconds, and hundreds of sources rather than a handful.[^deep-research] For a narrow, self-contained question, plain semantic search finishes in one round trip — deep research is for the answer you cannot reach in one step.
0 commit comments