Give the agent layered context
閱讀約需 103 分鐘 分鐘,9 節 中共有 9 個範例。
此部分尚未翻譯,因此以英文呈現。
By the end of this chapter, you can give a model that writes SQL more than the schema, in layers: notes on columns whose names mislead, a glossary of the shop's business terms, and verified queries retrieved for each question. You can fit those layers to a token budget, measure what each one adds on the dev split of the golden questions, and check the best set once on the held-out test split.
A schema can't say what its columns mean. Chapter 1 gave a model only the shop's schema and asked for net revenue in each region in March 2026. The query it wrote ran, looked right and counted cancelled orders. The rule it missed exists, in the finance team's definition of net revenue, but nothing the model was given contains it. Layered context is the most direct fix for what a schema can't say: write the rules down where the model reads them. It isn't the fix this book ends with, since chapters 5 and 6 move the definitions into a semantic layer the model can't misread. But any agent that writes SQL needs context, including the guarded one in chapter 7, and the habit this chapter teaches, adding one layer at a time and measuring each, applies to every change that follows.
此部分尚未翻譯,因此以英文呈現。
What the schema says, and what it doesn't
A schema tells a model which tables exist, what their columns are called and what type each column holds. That is a lot, and it never goes out of date, because the database produces it. What it can't tell the model is what the columns mean: which rows a table holds, which rows count toward a business number, which time zone a timestamp is in, and which of two similar tables is the one people trust. Those meanings decide whether a query returns the right number, and they live in the SQL that built the marts and in the heads of the people who own them.
The book's measure of this is the golden questions, which chapter 3 built: questions with a verified answer, scored on the dev split while context and prompts are tuned, and once on the held-out test split. Every layer in this chapter is added to the schema, so the schema is the first layer, and the place to start is what it leaves open. build_context() in lib/context.py, the shared module that builds every layer in this chapter, returns the schema layer when it's asked for nothing else.
1"""What the schema tells a model about revenue, and three March 2026 totals that all fit it. The three printed totals2carry the paragraph's point: they differ, and only the one that leaves out cancelled orders follows the finance definition."""3from lib.context import build_context4from lib.warehouse import connect56con = connect()7schema = build_context(con, "", layers=("schema",)) # no question, so only the schema layer is built8print(f"tables in the schema: {schema.text.count('CREATE TABLE')}")9print(f"schema tokens (estimated): {schema.tokens}")10for line in schema.text.splitlines():11 if "analytics.fct_orders " in line or "analytics.orders_legacy " in line:12 print(line)1314march = "BETWEEN DATE '2026-03-01' AND DATE '2026-03-31'" # the window of the question, on the business day15candidates = {16 "orders_legacy.revenue": f"SELECT sum(revenue) FROM analytics.orders_legacy WHERE order_date {march}",17 "fct_orders.net_revenue_eur, every order":18 f"SELECT sum(net_revenue_eur) FROM analytics.fct_orders WHERE order_date_local {march}",19 "fct_orders.net_revenue_eur, not cancelled":20 f"SELECT sum(net_revenue_eur) FROM analytics.fct_orders WHERE order_date_local {march} "21 "AND status <> 'cancelled'", # net revenue leaves out cancelled orders, as the status note says22}23for name, sql in candidates.items():24 (total,), = con.execute(sql).fetchall()25 print(f"{name}: {total:,.2f}")
python schema-digest.pytables in the schema: 16
schema tokens (estimated): 912
CREATE TABLE analytics.fct_orders (order_id INTEGER, customer_id INTEGER, region VARCHAR, country VARCHAR, channel VARCHAR, status VARCHAR, placed_at TIMESTAMP, order_date_local DATE, currency VARCHAR, line_count BIGINT, gross_revenue_eur DECIMAL(12,2), net_revenue_eur DECIMAL(12,2), discount_eur DECIMAL(12,2), tax_eur DECIMAL(12,2), _loaded_at TIMESTAMP);
CREATE TABLE analytics.orders_legacy (id INTEGER, cust_id INTEGER, order_date DATE, order_status VARCHAR, sales_region VARCHAR, revenue DECIMAL(12,2));
orders_legacy.revenue: 186,635.93
fct_orders.net_revenue_eur, every order: 162,542.79
fct_orders.net_revenue_eur, not cancelled: 150,082.83
The schema layer is the whole warehouse: all 16 tables, about 912 tokens. Read the two printed lines as a model would. analytics.fct_orders has a net_revenue_eur column and a status column, and analytics.orders_legacy has a revenue column and an order_date.
Each of the three totals the example computes for March 2026 is a reasonable reading of those two lines, and they disagree. The legacy table's revenue gives 186,635.93 euros, and the mart's net_revenue_eur over every order gives 162,542.79, which is what chapter 1's model computed. Only the third, 150,082.83, follows the finance team's definition, and nothing in the schema points to it rather than the other two.
The whole of the shop's warehouse fits in those 912 tokens, so there's no case for trimming it: a study of strong models on the BIRD benchmark found that passing the whole schema when it fits can beat selecting tables first, because the selection step can drop a column the query needs (Maamari et al., 2024). A warehouse with thousands of tables is a different problem, which this chapter doesn't solve. The shop's problem isn't which tables to show. It's that what the tables show doesn't say which number is right.
Working rule: Keep the whole schema in the context when it fits, and add meaning to it rather than taking tables away. Before you write any other layer, take one real question and list every reading of the schema that answers it. Each reading the schema doesn't rule out is a rule some other layer has to state.
此部分尚未翻譯,因此以英文呈現。
Describe columns where names mislead
The schema leaves open which of several readings is right, and the cheapest place to close some of them is the column itself. A column note is a sentence about one table or column that says what its name doesn't: which rows the table holds, the unit and sign of an amount, the time zone of a timestamp, the values a column takes, or that a table is deprecated. A note isn't documentation of every column. A column called product_name needs none, and a note that repeats the name costs tokens and says nothing. The notes worth writing are on the columns a careful analyst new to the warehouse would misread.
Published evidence points the same way. A study that generated column descriptions with a language model found that adding them to the prompt improved text-to-SQL accuracy and, in an ablation that replaced column names with uninformative ones, that the descriptions helped most there (Wretblad et al., 2024); it is a single preprint on the BIRD benchmark. Databricks' guidance for its Genie Agents says that "quality table and column descriptions in Unity Catalog are critical for Genie accuracy" (Databricks), which is the vendor's advice, not a measurement.
The shop's notes live in examples/data/shop/columns.toml, one entry per table with an optional note per column, and load_column_notes() in lib/context.py turns each into one line, such as analytics.fct_orders.placed_at: When the order was placed, in UTC. The notes cover the traps chapter 2 found: orders_legacy is deprecated, net_revenue_eur still holds a value for cancelled orders, placed_at is UTC while order_date_local is the business day, discount is negative in raw.order_lines and positive in the marts, and the fiscal year starts on February 1.
The recording context-layers shows what the notes do once the examples replay it. The recording holds the pinned model's SQL for every dev question of kind answer, written once with each layer set this chapter measures, using the prompt LAYERED_SQL in lib/prompts.py. The recording was made by recordings/record_context_layers.py, with the same model, sampling and seed as chapter 1's, and the examples replay it by the hash of the exact messages.
1"""Add the column notes on columns whose names mislead, and count the dev questions they fix and break under the recorded2model. The `fixed` and `broken` lines settle the net effect of the notes, and the broken question shows how a note misleads."""3from lib.context import LAYER_SETS, build_context, grade_with_context, load_column_notes4from lib.golden import load_golden5from lib.model import RecordedModel6from lib.warehouse import connect78con = connect()9notes = load_column_notes()10print(f"column notes: {len(notes)}")11print(f"column notes tokens (estimated): {build_context(con, '', layers=('columns',)).tokens}")12for note in notes:13 if note.startswith(("analytics.fct_orders.status", "analytics.fct_orders.placed_at")):14 print(note)1516model = RecordedModel("context-layers")17before, after = LAYER_SETS[0], LAYER_SETS[1] # the sets are cumulative: [0] is the schema alone, [1] adds the notes18dev = [q for q in load_golden() if q.split == "dev" and q.kind == "answer"]19fixed, broken = [], []20for q in dev:21 was, now = grade_with_context(con, model, q, before), grade_with_context(con, model, q, after)22 if was != "correct" and now == "correct": # fixed: wrong or error without the notes, correct with them23 fixed.append(q.question_id)24 if was == "correct" and now != "correct":25 broken.append(q.question_id)26print(f"fixed by column notes: {len(fixed)} ({', '.join(fixed)})")27print(f"broken by column notes: {len(broken)} ({', '.join(broken)})")
python column-descriptions.pycolumn notes: 33
column notes tokens (estimated): 1087
analytics.fct_orders.status: One of placed, shipped, delivered, cancelled, returned. Revenue leaves out cancelled orders; order and customer counts include them. Returned orders count everywhere.
analytics.fct_orders.placed_at: When the order was placed, in UTC. Don't filter or group dates on it; use order_date_local.
fixed by column notes: 7 (gq-001, gq-002, gq-003, gq-004, gq-008, gq-009, gq-011)
broken by column notes: 1 (gq-018)
The 33 notes add about 1087 tokens, more than the schema itself. On the dev questions they fix 7 (gq-001, gq-002, gq-003, gq-004, gq-008, gq-009, gq-011), among them gq-002, the March question from chapter 1, whose query now filters out cancelled orders. They also break 1 (gq-018). That question asks for the refunds on orders placed in March 2026, and with the notes the model filtered on both the order's date and refund_month, so it dropped refunds paid in a later month. The note on refund_month was accurate. The model used it on a question that didn't call for it, which is what a note can't prevent.
Working rule: Write notes only for tables and columns whose names mislead: deprecated tables, amounts with a sign or unit you wouldn't guess, timestamps in a time zone other than the business's, and columns whose values include a placeholder such as (none). Write them as facts about the column, not as instructions for one question. Expect a note to fix many questions and occasionally mislead on one, and keep the questions it broke in the golden set so you see it happen.
此部分尚未翻譯,因此以英文呈現。
Add the business glossary
Column notes describe the tables. Many questions turn on something else: what a business word means at this company. "Revenue" could be gross or net. "Orders" could include cancelled ones or not. Refunds "in March" could mean refunds paid in March or refunds on March's orders. These are choices, made by the teams that own the numbers, and no table states them. A business glossary lists the business words whose meaning a company's teams chose, and writes each meaning in terms of the warehouse's tables and columns.
Evidence for this layer comes from BIRD, a text-to-SQL benchmark whose questions come with "external knowledge evidence": hints that link a question's words to the database, such as what a term means as a formula over columns. In the benchmark's paper, adding the evidence raised GPT-4's execution accuracy on the dev set from 30.90% to 46.35%, with 2023 models and zero-shot prompts (Li et al., 2023). The hints are written per question, though, which is closer to giving the model the answer's ingredients than to giving it a glossary written once.
The shop's glossary is examples/data/shop/glossary.toml. Its entries define revenue as net revenue unless the question says gross, orders as every order including cancelled ones, average order value, fiscal periods, new customers, refunds by the month they were paid, units sold, and the two meanings of "channel". load_glossary() turns each into one line.
The glossary wasn't written blind, and the book's history shows how it changed. In the same change that recorded the model's first dev replies, with the schema alone, the entry for orders was corrected: it had said that order counts leave out cancelled orders, which disagreed with the finance team's golden SQL. The entries for units sold and for customers who ordered were added in that change too, and the column note on status was reworded to match. Later, after the test questions had been recorded with the schema alone, the refund-rate entry was changed to the book-wide definition that chapter 5 uses and that test question gq-028 also follows: refunded amount divided by gross revenue, because a refund pays back tax and only gross revenue includes it. Each change made the glossary agree with golden SQL that it was then scored against. That is how a dev score gets fitted to dev, and the check on test comes back to what it costs.
1"""Add the business glossary on top of the column notes, and count what it changes on the dev questions. The gq-0022lines show the query the prose describes: with the glossary, the March query no longer filters on status."""3from lib.context import LAYER_SETS, build_context, grade_with_context, load_glossary4from lib.golden import load_golden5from lib.model import RecordedModel6from lib.prompts import layered_sql_messages7from lib.warehouse import connect89con = connect()10terms = load_glossary()11print(f"glossary terms: {len(terms)}")12print(f"glossary tokens (estimated): {build_context(con, '', layers=('glossary',)).tokens}")13print(next(t for t in terms if t.startswith("revenue:")))1415model = RecordedModel("context-layers")16before, after = LAYER_SETS[1], LAYER_SETS[2] # schema and column notes; then the glossary too17dev = [q for q in load_golden() if q.split == "dev" and q.kind == "answer"]18changes = {(was, now): [] for was in ("correct", "other") for now in ("correct", "other")} # "other" is wrong or error19for q in dev:20 was, now = (grade_with_context(con, model, q, layers) for layers in (before, after))21 changes["correct" if was == "correct" else "other", "correct" if now == "correct" else "other"].append(q.question_id)22print(f"fixed by the glossary: {len(changes['other', 'correct'])} ({', '.join(changes['other', 'correct'])})")23print(f"broken by the glossary: {len(changes['correct', 'other'])} ({', '.join(changes['correct', 'other'])})")2425q = next(q for q in dev if q.question_id == "gq-002")26reply = model.complete(layered_sql_messages(build_context(con, q.question, after).text, q.question))27filters_status = "status" in reply # a query that filters on status has to name the column28assert not filters_status # the chapter states it: with the glossary, gq-002 no longer filters on status29print(f"gq-002 with the glossary filters status: {'yes' if filters_status else 'no'}")
python glossary-context.pyglossary terms: 13
glossary tokens (estimated): 609
revenue: Net revenue unless the question says gross: sum of analytics.fct_orders.net_revenue_eur over orders whose status isn't 'cancelled'. In EUR.
fixed by the glossary: 3 (gq-012, gq-015, gq-018)
broken by the glossary: 3 (gq-002, gq-009, gq-016)
gq-002 with the glossary filters status: no
The glossary's gains and losses cancel out: it fixes 3 (gq-012, gq-015, gq-018) dev questions and breaks 3 (gq-002, gq-009, gq-016). Its 13 terms add about 609 tokens on top of the schema and the column notes. The broken ones are the instructive part. For gq-002, the same March question, the query written with the glossary no longer filters on status, although the revenue entry and the note on status both say to leave cancelled orders out. For gq-016, the model left cancelled orders out of a customer count, applying the revenue rule where the glossary says it doesn't apply. For gq-009, the average order value, it divided the average order's net revenue by the number of orders again, although the glossary defines the measure as net revenue divided by the number of orders.
Each configuration was recorded once, with one sampled reply per question, so a change of one or two questions out of 17 is within what a second sample could undo. Chapter 10 measures that spread with repeated runs. What this result supports is narrower: the glossary didn't make this model better on these questions, and a rule stated twice was still missed. A plausible reason, which we haven't measured here, is that a longer context makes a small model use each part of it less reliably. Studies of long contexts find that models use information in the middle of a prompt worse than at its edges (Liu et al., 2024), and that accuracy falls with length when the question and the passage don't share words (Modarressi et al., 2025).
Working rule: Write a glossary entry for each business word whose meaning is a choice someone made, and state it in terms of tables and columns. Prefer one place for each rule, a column note or a glossary entry; the shop's files state the cancelled-orders rule in both, and the miss on gq-002 shows that saying it twice doesn't make it stick. Measure the glossary on your dev questions before you trust it, and keep it only if it helps there or the next layer depends on it.
此部分尚未翻譯,因此以英文呈現。
Retrieve verified queries as examples
Column notes and a glossary state rules. A verified query shows the shop's rules applied: a question someone asked, with SQL a person checked against the shop's definitions. Given a few such pairs that resemble the question, a model can copy their shape, such as the join to dim_date for a fiscal period or the filter on status, instead of inventing one. A systematic comparison of prompt designs for text-to-SQL found that the best setup chose examples by the similarity of their questions, after masking domain-specific words, and by the similarity of their SQL to a draft query written first, and showed each as a question and its SQL alone (Gao et al., 2023). Warehouse vendors build the same idea into their assistants. Snowflake's Cortex Analyst keeps a verified query repository of "questions and corresponding SQL queries to answer them" (Snowflake), and Databricks says a Genie Agent can use an example SQL query directly or learn from it, and treats parameterized ones as trusted assets (Databricks). Both are the vendors' own descriptions, not measurements.
The shop's verified queries live in examples/data/shop/verified_queries.toml. Their fields follow Snowflake's field names, name, question, sql, verified_at and verified_by, plus the role the query was checked for. Two differ in content. Snowflake's verified_at is optional and counted in seconds since 1970, and its verified_by names the person who checked the query; the shop's files give a date and name a team, never a person. None of the verified queries is a golden question. They are near neighbours of questions in both splits, such as net revenue by fiscal quarter for fiscal 2025 or average order value by region for 2024, which is how a real repository grows: from questions people asked before. A verified query that was itself a golden question would make the score measure copying.
Retrieval in lib/context.py is lexical, with no embeddings. question_terms() lowercases the question, drops common words and masks three kinds of word: numbers, month names and the shop's three region codes. Every other word stays as it is. QueryIndex scores each verified question against the masked question with TF-IDF and cosine similarity, and search() keeps at most three matches that score at least 0.4. This is a cheap stand-in for the selection in Gao et al., not a reproduction of it: it masks far fewer words, and it drops the second step, which compares draft SQL.
1"""Retrieve the verified queries most like a question, by TF-IDF over the masked question's words. The matches kept at the20.4 default show that a threshold can leave a question with no example, and the masked phrases show what masking does."""3from lib.context import QueryIndex, load_verified_queries, question_terms45queries = load_verified_queries()6index = QueryIndex(queries)7print(f"verified queries: {len(queries)}")89for label, question in (("aov", "What was the average order value in each region in March 2026?"),10 ("parcels", "Which warehouse ships the most parcels?")):11 print(f"question: {question}")12 print(f"masked terms: {' '.join(question_terms(question))}")13 for score, query in index.search(question, k=3, min_score=0.0): # no threshold here, so the scores show the gap14 print(f" {score:.3f} {query.name}")15 print(f"{label} matches kept: {len(index.search(question))}") # the default threshold, 0.4, is the one the book uses1617# Months and numbers are masked; "calendar" and "year" are not common words, and "Q3" is not masked at all.18for phrase in ("March 2026", "January 2025", "calendar year 2024", "EMEA in Q3"):19 print(f"masked \"{phrase}\": {' '.join(question_terms(phrase))}")
python retrieve-similar-queries.pyverified queries: 16
question: What was the average order value in each region in March 2026?
masked terms: average order value region <month> <number>
0.835 average_order_value_by_region
0.422 gross_revenue_by_region_month
0.191 orders_of_one_customer
aov matches kept: 2
question: Which warehouse ships the most parcels?
masked terms: warehouse ships most parcels
0.375 units_sold_by_product
0.357 net_revenue_by_category
0.000 average_order_value_by_region
parcels matches kept: 0
masked "March 2026": <month> <number>
masked "January 2025": <month> <number>
masked "calendar year 2024": calendar year <number>
masked "EMEA in Q3": <region> q3
For the average order value in each region in March 2026, the closest of the 16 verified queries is the same measure by region for 2024, and 2 matches pass the threshold. The second question, about which warehouse ships the most parcels, is one the shop's data can't answer at all. Its best scores come from the single shared word "most", and none passes the threshold, so 0 queries are offered. Without the threshold, the model would get three confident-looking examples, two of them about products, for a question about shipping.
The last lines of the output show what masking does and doesn't do. "March 2026" and "January 2025" both become <month> <number>, so a question about one month matches a verified query about another. "Calendar year 2024" becomes calendar year <number>, because "calendar" and "year" are neither common words nor values, and a quarter written as "Q3" isn't masked at all. That is why the March question's best match, a question about calendar year 2024, scores less than a perfect 1.
Working rule: Retrieve a few verified queries per question, chosen by similarity of the masked question, and show each as its question and SQL. Set a minimum score and check it on questions the warehouse can't answer, because lexical retrieval always returns something. Never let a golden question into the verified set. Move to embeddings only when your own measurement shows lexical retrieval missing queries that a person would call similar.
讓上下文符合預算
有了結構描述(schema)、欄位說明、詞彙表和檢索到的已驗證查詢,商店的上下文所佔的詞元數已是單用結構描述時的數倍,而且每多加一層,上下文就更長。詞元預算是代理程式針對一個問題最多可以送出的上下文量。它的存在有三個理由。第一,模型的視窗有限:本書錄製時使用的執行環境給模型 8,192 個詞元,其中最多 384 個保留給回覆。第二,每個提示詞元在 CPU 上耗費時間,使用託管模型時則耗費金錢。第三,較長的上下文不會被平均利用,因為模型對長提示中間資訊的利用程度,低於對開頭或結尾資訊的利用程度(Liu et al., 2024)。
讓上下文符合預算,就是決定要捨棄什麼,而這個決定需要一套優先順序。build_context() 依優先順序取用各層,逐項加入其中的項目,並捨棄任何會使上下文超出預算的項目,因此較後面、較短的項目仍可能放得下。結構描述絕不會被捨棄,原因見第一節所述:放得下的結構描述應完整送出。無論優先順序為何,上下文都以固定順序呈現:穩定的層在前,檢索到的查詢接在後面,問題放在最後。這樣可以讓問題緊鄰與它最相似的範例,也讓模型執行器能重複使用每個問題共有的那部分提示,而不必在每次呼叫時重新讀取。製作本書錄製內容的執行器 llama-cpp-python 0.3.36 在 Llama.generate 中做了這件事:它把新提示的詞元與上一個提示的詞元相比,只評估兩者最長共同前綴之後的詞元(llama-cpp-python 0.3.36,llama.py)。
lib/context.py 中的詞元數是估計值:estimate_tokens() 把字元數除以 3.5,因為這些範例不載入模型的分詞器。下面把這個估計值與錄製提示時模型分詞器回報的計數相比較,並針對平均訂單金額的問題 gq-009,在 2,400 個詞元的示範預算下,以兩種優先順序來調整其上下文。
1"""Fit one question's layered context to a token budget, and check the token estimate against the model's count. The2estimate runs high, the safe side for a budget, and the dropped lines show what each priority gives up."""3from lib.context import LAYERS, build_context, estimate_tokens4from lib.golden import load_golden5from lib.model import RecordedModel6from lib.prompts import layered_sql_messages7from lib.warehouse import connect89con = connect()10question = next(q for q in load_golden() if q.question_id == "gq-009").question11full = build_context(con, question)12for layer, tokens in full.layer_tokens().items():13 print(f"{layer} tokens (estimated): {tokens}")1415messages = layered_sql_messages(full.text, question)16call = RecordedModel("context-layers").call_for(messages)17print(f"prompt tokens (estimated): {estimate_tokens(''.join(m['content'] for m in messages))}") # every message, joined18print(f"prompt tokens (model's tokenizer): {call['prompt_tokens']}")1920budget = 2400 # an illustrative budget, well inside the 8,192-token window the recordings use21for name, priority in (("default", LAYERS), ("examples first", ("schema", "columns", "verified", "glossary"))): # verified ranks above glossary22 fitted = build_context(con, question, layers=priority, budget=budget)23 kept = {layer: len(items) for layer, items in fitted.items.items()}24 print(f"{name} priority: {fitted.tokens} tokens, kept {kept}")25 print(f"{name} dropped: {len(fitted.dropped)}")26print(f"render order: {', '.join(layer for layer, items in fitted.items.items() if items)}, then the question")
python context-token-budget.pyschema tokens (estimated): 912
columns tokens (estimated): 1087
glossary tokens (estimated): 609
verified tokens (estimated): 200
prompt tokens (estimated): 2917
prompt tokens (model's tokenizer): 2364
default priority: 2393 tokens, kept {'schema': 1, 'columns': 33, 'glossary': 9, 'verified': 0}
default dropped: 6
examples first priority: 2384 tokens, kept {'schema': 1, 'columns': 33, 'glossary': 5, 'verified': 2}
examples first dropped: 8
render order: schema, columns, glossary, verified, then the question
整個提示的估計值為 2917 個詞元,而模型分詞器計得的數字為 2364。估計偏高對預算而言是安全的方向:它所接納的上下文一定放得下。在預設優先順序下(依各層被介紹的順序),預算保留所有欄位說明和部分詞彙表,並捨棄兩筆檢索到的查詢,共捨棄 6 項。若把已驗證查詢的順序排在詞彙表之上,同樣的預算會保留這兩筆查詢,改為捨棄 8 筆詞彙表項目。兩種順序都不是普遍正確的。下一節會衡量哪些層在商店的問題上值得佔用詞元,而這才是決定順序的依據。
實用原則: 把詞元預算設定在遠低於模型視窗之處,並依你衡量過的優先順序,而不是依撰寫各層的順序,讓每個問題的上下文符合預算。結構描述放得下時,絕不要捨棄。無論捨棄了什麼,都把穩定的層放在前面、問題放在最後。把你的詞元估計值與模型的實際計數核對一次,並讓估計值維持在偏高的一側。
此部分尚未翻譯,因此以英文呈現。
Measure what each layer adds
Each layer so far was judged by the questions it fixed and broke. To choose which layers to keep, and in what order a budget should drop them, you need all of them measured the same way. An ablation measures every layer the same way: it scores a series of configurations that differ by one layer each, on the same questions, under the same comparison rule. The questions are the dev split of the golden questions, because choosing among configurations is tuning, and tuning on the test split would leave nothing held out to check the choice.
The shop's dev split has 24 golden questions. This chapter scores the 17 of kind answer. The other 7 should be declined, because the warehouse can't answer them or the role asking may not see the answer, and a prompt that only writes SQL can't decline, so they would score as wrong under every layer set; chapters 8 and 9 score them. Each layer set adds one layer to the one before: the schema alone, then column notes, then the glossary, then up to three retrieved verified queries. The outcomes come from grade() in lib/golden.py, as in every chapter: correct when the result matches the golden result, wrong when it runs and doesn't, error when it doesn't run. The first layer set, the schema alone, isn't chapter 3's baseline either. That baseline used the schema-only prompt and scored every question in each split, while this run uses the layered prompt LAYERED_SQL with no other layer filled in and scores the answer questions only, so its numbers differ.
1"""Score each layer set on the dev split's answer questions, replaying the recorded model's SQL for each, and print the2totals, the average token counts and the outcome grid. The correct totals and the fixed and broken lines are the numbers the prose rests on."""3from collections import Counter45from lib.context import LAYER_SETS, build_context, grade_with_context, layer_set_name6from lib.golden import load_golden7from lib.model import RecordedModel8from lib.warehouse import connect910con = connect()11model = RecordedModel("context-layers")12dev_split = [q for q in load_golden() if q.split == "dev"]13dev = [q for q in dev_split if q.kind == "answer"] # the other questions are declined, and are scored in chapters 8 and 914print(f"dev split questions: {len(dev_split)}")15print(f"dev questions: {len(dev)}")16print(f"dev questions to decline: {len(dev_split) - len(dev)}")1718grid = {}19for layers in LAYER_SETS:20 name = layer_set_name(layers)21 grid[name] = [grade_with_context(con, model, q, layers) for q in dev]22 counts = Counter(grid[name])23 tokens = sum(build_context(con, q.question, layers).tokens for q in dev) // len(dev) # mean over the dev questions, whole tokens24 print(f"{name}: correct {counts['correct']}, wrong {counts['wrong']}, error {counts['error']}")25 print(f"{name} correct: {counts['correct']}")26 print(f"{name} error: {counts['error']}")27 print(f"{name} tokens: {tokens}")2829print("question " + " ".join(f"{i + 1}" for i in range(len(LAYER_SETS))))30for i, q in enumerate(dev):31 print(f"{q.question_id} " + " ".join(grid[name][i][0] for name in grid))3233# What the verified queries changed, against the run with the glossary.34before, after = grid[layer_set_name(LAYER_SETS[2])], grid[layer_set_name(LAYER_SETS[3])] # the glossary run, then the glossary plus verified queries35fixed = [q.question_id for q, was, now in zip(dev, before, after) if was != "correct" and now == "correct"]36broken = [q.question_id for q, was, now in zip(dev, before, after) if was == "correct" and now != "correct"]37print(f"fixed by verified queries: {len(fixed)}")38print(f"fixed by verified queries, IDs: {', '.join(fixed)}")39print(f"broken by verified queries: {len(broken)}")
python layer-ablation-dev.pydev split questions: 24
dev questions: 17
dev questions to decline: 7
schema: correct 3, wrong 9, error 5
schema correct: 3
schema error: 5
schema tokens: 912
schema+columns: correct 9, wrong 7, error 1
schema+columns correct: 9
schema+columns error: 1
schema+columns tokens: 1999
schema+columns+glossary: correct 9, wrong 8, error 0
schema+columns+glossary correct: 9
schema+columns+glossary error: 0
schema+columns+glossary tokens: 2608
schema+columns+glossary+verified: correct 14, wrong 2, error 1
schema+columns+glossary+verified correct: 14
schema+columns+glossary+verified error: 1
schema+columns+glossary+verified tokens: 2775
question 1 2 3 4
gq-001 e c c c
gq-002 w c w c
gq-003 e c c c
gq-004 w c c c
gq-006 e e w w
gq-007 w w w w
gq-008 w c c c
gq-009 w c w c
gq-010 e w w c
gq-011 w c c c
gq-012 w w c c
gq-013 e w w c
gq-014 c c c c
gq-015 w w c c
gq-016 c c w c
gq-017 w w w e
gq-018 c w c c
fixed by verified queries: 5
fixed by verified queries, IDs: gq-002, gq-009, gq-010, gq-013, gq-016
broken by verified queries: 0
Column notes and verified queries raise the score; the glossary doesn't. The schema alone gets 3 of the dev questions right, column notes take it to 9, and the glossary leaves it at 9. Verified queries on top give 14, the best set on dev.
The grid shows where the change came from. Against the run with the glossary, retrieved examples fixed 5 questions and broke 0. Two of the fixes are new: the fiscal-quarter question gq-010 and the order count by region gq-013, whose nearest verified queries show exactly the join and the count they needed. The other three, gq-002, gq-009 and gq-016, are the questions the glossary run had lost, so there the examples won back what the glossary had cost.
The totals say how much each layer helped. The failures say what the next layer should be, so the next two examples read them, starting with the schema-only queries that don't run.
1"""Why the dev queries written with the schema alone don't run: the column each one names, and the analytics tables that have it.2The per-table counts are the evidence the prose rests on: every failure is a column on the wrong table."""3import re4from collections import Counter56import duckdb7from lib.context import LAYER_SETS, build_context8from lib.golden import load_golden9from lib.model import RecordedModel10from lib.prompts import layered_sql_messages11from lib.tools import extract_sql, run_sql12from lib.warehouse import connect1314con, model = connect(), RecordedModel("context-layers")15dev = [q for q in load_golden() if q.split == "dev" and q.kind == "answer"]16where = Counter()17for q in dev:18 context = build_context(con, q.question, LAYER_SETS[0]) # the schema alone19 sql = extract_sql(model.complete(layered_sql_messages(context.text, q.question)))20 try:21 run_sql(con, sql)22 continue23 except duckdb.Error as e:24 column = re.search(r'column (?:named )?"(\w+)"', str(e)).group(1) # DuckDB says either form; take the name25 (tables,), = con.execute(26 "SELECT string_agg(table_schema || '.' || table_name, ', ' ORDER BY 1) FROM information_schema.columns " # sorted, so a column in two tables lists them in a fixed order27 "WHERE table_schema = 'analytics' AND column_name = ?", [column]).fetchall() # the marts only: where the model should have looked28 where[tables] += 129 print(f"{q.question_id}: no column {column}; it is in {tables}")30print(f"schema-only queries that don't run: {where.total()}")31for tables, n in where.items():32 print(f"missing column only {tables} has: {n}")
python missing-columns-schema-only.pygq-001: no column fiscal_year; it is in analytics.dim_date
gq-003: no column calendar_year; it is in analytics.dim_date
gq-006: no column quantity; it is in analytics.fct_order_lines
gq-010: no column fiscal_quarter; it is in analytics.dim_date
gq-013: no column order_date; it is in analytics.orders_legacy
schema-only queries that don't run: 5
missing column only analytics.dim_date has: 3
missing column only analytics.fct_order_lines has: 1
missing column only analytics.orders_legacy has: 1
Every failure is a column the model put on the wrong table. Of the 5, 3 ask the orders mart for a fiscal or calendar column that only dim_date has, 1 asks it for quantity, which only the order lines have, and 1 uses the legacy table's order_date.
These are the mistakes column notes are for. With the notes, 1 query in the whole dev split fails to run.
A second script prints the query for each dev question that all four layers still don't get right.
1"""The dev questions that stay wrong with all four layers, each with its query, and a check of whether gq-007 fails only on2the spelling of 'tents'. The count line settles how many dev questions stay wrong, and the gq-007 line settles whether the spelling alone causes its failure."""3import duckdb4from lib.context import LAYERS, build_context, grade_with_context5from lib.golden import golden_result, grade, load_golden6from lib.model import RecordedModel7from lib.prompts import layered_sql_messages8from lib.tools import extract_sql, run_sql9from lib.warehouse import connect1011con, model = connect(), RecordedModel("context-layers")12dev = [q for q in load_golden() if q.split == "dev" and q.kind == "answer"]131415def sql_with_all_layers(q):16 return extract_sql(model.complete(layered_sql_messages(build_context(con, q.question, LAYERS).text, q.question)))171819unsolved = [(q, outcome) for q in dev if (outcome := grade_with_context(con, model, q, LAYERS)) != "correct"] # LAYERS is all four20print(f"not correct with all four layers: {len(unsolved)}")21for q, outcome in unsolved:22 sql = sql_with_all_layers(q)23 print(f"{q.question_id} ({outcome}): {q.question}")24 print("\n".join(" " + line for line in sql.strip().splitlines()))25 if outcome == "error":26 try:27 run_sql(con, sql)28 except duckdb.Error as e:29 print(f" -- {str(e).splitlines()[0]}")3031gq007 = next(q for q in dev if q.question_id == "gq-007")32respelled = sql_with_all_layers(gq007).replace("'tents'", "'Tents'") # change the spelling only, to see whether it is the whole failure33print(f"gq-007 with 'Tents' spelled as stored: {grade(gq007, run_sql(con, respelled), golden_result(con, gq007))}")
python what-stays-wrong.pynot correct with all four layers: 3
gq-006 (wrong): What were net revenue and the number of units sold in each region in 2025?
SELECT
o.region,
SUM(o.net_revenue_eur) AS net_revenue,
SUM(l.quantity) AS units_sold
FROM analytics.fct_orders AS o
JOIN analytics.fct_order_lines AS l USING (order_id)
WHERE o.order_date_local >= '2025-01-01' AND o.order_date_local < '2026-01-01'
AND o.status <> 'cancelled'
GROUP BY o.region
ORDER BY net_revenue DESC
gq-007 (wrong): How much net revenue came from orders that included a tent in 2025?
SELECT
SUM(fol.net_amount_eur) AS net_revenue
FROM
analytics.fct_order_lines fol
JOIN
analytics.fct_orders fo ON fol.order_id = fo.order_id
JOIN
analytics.dim_products dp ON fol.product_id = dp.product_id
JOIN
analytics.dim_date dd ON fo.order_date_local = dd.date_day
WHERE
dp.category = 'tents'
AND dd.calendar_year = 2025
AND fo.status != 'cancelled'
gq-017 (error): What was net revenue in each of our countries in 2025?
SELECT d.country, round(sum(o.net_revenue_eur), 2) AS net_revenue
FROM analytics.fct_orders AS o
JOIN analytics.dim_date AS d ON d.date_day = o.order_date_local
WHERE o.status <> 'cancelled'
AND d.calendar_year = 2025
GROUP BY d.country
ORDER BY net_revenue DESC
-- Binder Error: Table "d" does not have a column named "country"
gq-007 with 'Tents' spelled as stored: wrong
With all four layers, 3 questions are still not right. gq-006 asks for net revenue and units sold by region. The model joined the orders mart to its lines and summed net_revenue_eur over the joined rows, so an order's revenue counts once for every line it has. The glossary's entry for units sold says to sum the lines per order first for exactly this reason, so this is a rule the context stated and the model missed. gq-007 filters on the category 'tents' where the data says 'Tents', a value no layer lists. Spelled as stored, the answer is still wrong, because the query sums only the tent lines, where the question asks for the whole revenue of the orders that included one. gq-017 asks for "our countries", which depends on the region of the regional lead asking. No layer says who is asking, and the model's query, which takes country from dim_date, doesn't run.
This is the book's measurement on one synthetic warehouse and one 4-billion-parameter model, with one sampled reply per question, not a general result about layered context. We found no peer-reviewed study from 2025 or 2026 that measures retrieved verified queries separately from other context on business data, which is why the chapter measures it here.
Working rule: Measure context one layer at a time on the dev split, with the same questions, comparison rule and model, and keep the outcomes per question, not only the totals. Treat a difference of one or two questions on a set this size as noise until repeated runs say otherwise. Read the queries that fail, because each one names a layer you don't have yet or a rule the model still skips.
在測試集上檢查選定的層
消融實驗在開發集上選出了全部四層,開發集就是本章用來調整的那些問題。在開發集上得分最高的那一組,仍須在保留的測試集上檢查一次,並與僅使用綱要的設定並列比較,這樣比較結果才不會因為是用同一批問題選出來的而顯得偏高。錄製檔 context-layers-test 收錄了模型針對 17 個類型為 answer 的測試問題所產生的 SQL,且只涵蓋這兩種層組合。每個測試問題也會與其最相符的已驗證查詢比較並計分,供後續段落使用。
1"""Choose the best layer set on the dev split, then score it once on the held-out test split beside the schema alone. The2correct counts are the numbers the prose rests on, and the retrieval scores measure how close each split sits to the verified queries."""3from collections import Counter45from lib.context import LAYER_SETS, QueryIndex, grade_with_context, layer_set_name, load_verified_queries6from lib.golden import load_golden7from lib.model import RecordedModel8from lib.warehouse import connect910con = connect()11questions = [q for q in load_golden() if q.kind == "answer"] # the answer questions only, as in the ablation12dev = [q for q in questions if q.split == "dev"]13test = [q for q in questions if q.split == "test"]1415dev_model = RecordedModel("context-layers") # dev and test replies come from separate recordings; the test one plays no part in the choice16dev_correct = {layers: sum(grade_with_context(con, dev_model, q, layers) == "correct" for q in dev)17 for layers in LAYER_SETS}18best = max(LAYER_SETS, key=lambda layers: (dev_correct[layers], -len(layers))) # ties go to fewer layers19print(f"best on dev: {layer_set_name(best)}")20print(f"best on dev correct: {dev_correct[best]} of {len(dev)}")2122test_model = RecordedModel("context-layers-test")23print(f"test questions: {len(test)}")24print(f"test question IDs: {', '.join(q.question_id for q in test)}")25for name, layers in (("schema", LAYER_SETS[0]), ("best", best)):26 counts = Counter(grade_with_context(con, test_model, q, layers) for q in test)27 print(f"test {name}: correct {counts['correct']}, wrong {counts['wrong']}, error {counts['error']}")28 print(f"test {name} correct: {counts['correct']}")2930# How close each split's questions sit to the verified queries, by the retriever's own score.31index = QueryIndex(load_verified_queries())32for split, qs in (("dev", dev), ("test", test)):33 best_match = [(index.search(q.question, k=1, min_score=0.0)[0], q) for q in qs] # the best match, whatever its score34 print(f"{split} mean top retrieval score: {sum(score for (score, _), _ in best_match) / len(qs):.2f}")35 print(f"{split} questions with no verified query kept: {sum(not index.search(q.question) for q in qs)}")36 (score, query), q = max(best_match, key=lambda pair: pair[0][0])37 print(f"{split} closest pair: {q.question_id} and {query.name}")38 print(f"{split} closest score: {score:.3f}")
python best-layers-on-test.pybest on dev: schema+columns+glossary+verified
best on dev correct: 14 of 17
test questions: 17
test question IDs: gq-025, gq-026, gq-027, gq-028, gq-029, gq-030, gq-031, gq-032, gq-033, gq-034, gq-035, gq-036, gq-037, gq-038, gq-039, gq-040, gq-041
test schema: correct 4, wrong 9, error 4
test schema correct: 4
test best: correct 8, wrong 7, error 2
test best correct: 8
dev mean top retrieval score: 0.57
dev questions with no verified query kept: 2
dev closest pair: gq-001 and net_revenue_by_fiscal_quarter
dev closest score: 0.862
test mean top retrieval score: 0.63
test questions with no verified query kept: 1
test closest pair: gq-029 and new_customers_by_region
test closest score: 0.933
這些層對沒有人針對它們調整過的問題仍然有幫助,而這才是重要的結果。開發集上最佳的組合是 schema+columns+glossary+verified,在開發集上答對 14 of 17。在測試集上,僅使用綱要答對 4,四層全用則答對 8。
開發集與測試集分數之間的差距是另一項結果,本章並未衡量其成因。其中一種解釋是已驗證查詢比較適合開發集,但檢索器自己的分數與此相反:每個問題最佳匹配的平均分數,在開發集上是 0.57,在測試集上是 0.63。
最接近的一組配對,也就是一個黃金問題與其已驗證的匹配查詢,同樣出現在測試集:gq-029 and new_customers_by_region,分數為 0.933,內容是對 2024 日曆年提出的同一個問題。開發集中最接近的一組得分為 0.862。
有兩點是已知的。第一,開發集的用途不只是選出組合。如詞彙表一節所述,在錄下最初幾筆開發集回覆時,詞彙表曾被修改以便與黃金 SQL 一致,這類修改會使上下文貼合它將被評分的那些問題,無論是否出於本意。開發集就是這樣被過度擬合的,這也是開發集分數不該作為回報結果的原因。第二,兩個資料集所問的內容不同:測試集詢問毛利、工作階段與行銷支出,開發集的問題完全沒有涉及這些。要區分這些成因,需要比本章更多的問題與重複執行。
實用原則: 在開發集上選定組合,然後在測試集上與基準線並列,只評分一次,並回報測試集的數字。請記錄在讀過開發集結果之後對上下文所做的每一項修改,因為每一項都可能使開發集分數被高估。當開發集與測試集的結果不一致時,先檢查可以衡量的事項,例如各資料集與你的範例有多接近,再去解釋差距。
此部分尚未翻譯,因此以英文呈現。
What goes wrong in production
The layers raised the dev score well above the schema alone, and they stay correct only while the warehouse and its definitions stay put. Once people rely on them, the failures are quiet: a note or a verified query keeps teaching a rule its warehouse has since changed.
The notes drift away from the marts. A column note is true about the mart as it was built when someone wrote it. When the analytics engineers change fct_orders, for example to drop cancelled orders from the mart, the note that says to filter on status becomes wrong, and nothing fails. You notice by keeping the notes in the same repository as the SQL that builds the marts, reviewing them in the same pull requests, and checking in CI that every table and column a note names still exists.
A value the context never lists. The model wrote 'tents' for a category stored as 'Tents', and the query ran and returned nothing. Questions name values, such as categories, channels, regions and campaign names, and a model guesses their spelling. Reading the wrong answers whose result is empty or zero catches it. For columns with a few dozen distinct values, list them in the column note. For columns with many, retrieve the values that match the question's words, as you retrieve verified queries.
A rule stated in context is still ignored. With the glossary, the model dropped the cancelled-orders filter on gq-002 and left cancelled orders out of the customer count for gq-016. With the column notes, and again with the glossary, it left them out of the order count for gq-013. Both go against what the note on status and the glossary say. With all four layers, it summed order revenue over order lines on gq-006, which the glossary's entry for units sold warns against. Context makes a rule available. It doesn't make the model follow it. Only per-question golden scores reveal it. When a rule has to hold every time, move it out of the prompt and into code the model can't skip, which is what chapter 5's semantic layer does.
Retrieval offers the wrong example. Lexical retrieval matches words, so a question that shares only "most" with a verified query gets that query as its example unless a threshold stops it, and a question that means the same thing in other words gets nothing. Log which verified query each answer used, as Snowflake's Cortex Analyst reports it in the confidence field of its API (Snowflake), then read the logs for answers that went wrong.
A verified query is mistaken for a verified answer. A retrieved query guides the model, and the SQL it writes is new SQL that nobody checked. Databricks marks a Genie answer as verified only when the exact text of a parameterized example query is used (Databricks), which is the distinction to keep. Never show an answer as verified because a similar query was. You notice by checking what your interface says about answers that were built from examples.
Verified queries go stale. A query checked in October uses the definitions of October. When finance changes a definition, the old queries teach the old rule to every similar question. Keep a verified_at date on each query, re-run each against its golden counterpart when a definition changes, and remove queries nobody has checked since the change.
The context grows until it costs more than it gives. Each layer is cheap to add, and nobody removes one that stopped helping. On a CPU every prompt token costs time, and with a hosted model it costs money on every call. Recording the tokens of each call with its outcome, as the recordings here do, and re-running the ablation when a layer grows, shows which layer has stopped paying for its tokens.
Working rule: Version the context with the warehouse: keep it beside the marts, checked in CI against the tables it names, scored on golden questions with every change, and logged per answer with the verified queries it used. Move any rule that must always hold out of the context and into a definition the model can't skip.
此部分尚未翻譯,因此以英文呈現。
Decisions this chapter made
- Context comes in four layers: the whole schema, notes on columns whose names mislead, a business glossary, and verified queries retrieved per question. Each lives in a file under
examples/data/shop/, andbuild_context()inlib/context.pyassembles them. - The schema is never trimmed while it fits; the budget applies to the other layers.
- Verified queries are retrieved lexically, by TF-IDF over the question's words with numbers, months and region codes masked, at most three, each with a score of at least 0.4. None of them is a golden question, and each names the team that verified it, never a person.
- The context is rendered stable layers first and the question last, so the question sits beside its examples and the runner can reuse the shared prefix.
- Tokens are estimated from characters, at 3.5 characters a token, which overestimated the model's count on the prompt checked, so a context that fits the estimate should fit the model; check it on your own prompts.
- Layers are measured by ablation on the dev split's answer questions, one layer at a time, and the best set is checked once on the test split beside the schema alone.
- The book keeps all four layers for the chapters that write SQL, because that set scored best on dev. The glossary stays although it didn't improve the score by itself: the verified-query layer was measured on top of it, and we didn't measure verified queries without it.
本章,於本版次中
版次 2026.11.15範例於 2026年11 月 19 通過測試本章 9 個範例 個範例中有 9 個通過
譯自英文書,版次 2026.11.15。