Extract Knowledge

Extract Business Knowledge from Gold SQL Pairs

You receive one or more (question, gold_sql) pairs. For each pair you must:

1. Produce a SQL candidate whose result set matches gold_sql exactly — either by simulating the drafting yourself (lite mode) or by driving a blind gen_sql subagent (deep mode). In deep mode, never expose the gold SQL or any gold result values to the subagent.
2. Diff the candidate against gold_sql to surface the business knowledge gap — the rules, joins, filters, granularity, or business definitions that a generic SQL agent would not know.
3. Filter the gap through the "worth-writing" test, then persist each surviving atom as the shortest possible fact into ./knowledge/<domain-slug>.md, and refresh the ## Knowledge index in ./AGENTS.md.

The goal is the shortest description that lets an LLM answer correctly. Matching the SQL is only the means; reusable atomic facts are the deliverable.

Critical Rules

• Never expose gold_sql to the subagent. Not the full SQL, not snippets, not concrete result values, not column-level data. The subagent must stay "blind" — only then do its mistakes reveal what knowledge is missing.
• Never quote gold's concrete values when talking to the subagent. Follow-up prompts may only describe symptoms and qualitative directions.
• Across retries, always reuse the subagent's session_id via the task tool — that's the only way the subagent remembers prior attempts.

Input

Resolve the pair source in this priority order (use the first one that works, do not keep looking):

1. Explicitly supplied in the current user message — single pair, or multiple (list, CSV file path, YAML/JSON).
2. Files / paths attached to the current message — if the user passes a path, read_file first, then parse.
3. Recovered from recent conversation context — when the user invokes /extract-knowledge without explicit pairs:
   • Scan backward from the latest message and locate the most recent user-approved (question, SQL) — i.e. a data question the user asked, followed by an assistant SQL the user continued to refine, accepted, or did not reject.
   • Use the user's original question as question, and the final adopted SQL as gold_sql.
   • Confirm via ask_user by showing the recovered question and a prefix of gold_sql. Do not run silently — context recovery has ambiguity risk (you may have latched onto an intermediate draft).

If none of the above yields a clear (question, SQL), call ask_user to request the pair from the user.

Mode Selection (lite / deep)

Once the input is resolved, decide which mode to run before entering the workflow:

• lite (default) — do not call the gen_sql subagent. The main agent itself simulates "given this question + the current datasource schema, how would I draft the SQL," diffs that mental draft against gold_sql, and mines the gap directly. Pros: fast, zero subagent calls, single-pass. Cons: no independent blind-generator validation; depends on the main agent's discipline to "pretend not to know the answer."
• deep — full pipeline. Drive the gen_sql subagent through multi-round blind iteration (≤5 rounds) until result match, then mine facts from the final diff. Pros: independent blind validator, iteration surfaces finer-grained boundary facts. Cons: consumes subagent tokens, requires multiple execute_sql executions.

Decision rule:

1. When the ask_user tool is available, call it once and ask the user to choose lite or deep; include the one-line tradeoff. Default suggestion = lite.
2. When ask_user is unavailable (no such tool, or running non-interactively), default to lite and note in the final output: "defaulted to lite; rerun and choose deep for stricter validation."

Record the chosen mode and route to the right workflow branch.

Workflow (one pair at a time)

Step 1 — Validate gold_sql (both modes)

Run execute_sql(sql=<gold_sql>).

• If it errors: report to the user and skip this pair. Do not invent questions to salvage a broken gold.
• Otherwise: cache the result (row count, column names, small preview). This is the factual baseline for later comparison.

Steps 2–4 (lite mode) — Main agent simulates and diffs

Take this branch only when mode = lite; skip the deep branch entirely.

1. Temporarily ignore gold_sql. Mentally switch to a state where you only know question + the current datasource schema. Do not let the specific shape of gold_sql influence this draft.
2. Write a draft SQL (no need to execute_sql-execute it; the draft is only a comparison baseline) — this is the version a SQL agent unfamiliar with this project's business conventions would produce given that schema.
3. Diff the draft against gold_sql along these axes:
   • Table choice, table aliases, missing mapping tables
   • Join type and join keys
   • WHERE filters (especially constant filters, status codes, tenant / platform predicates)
   • Group-by granularity, aggregate functions, deduplication strategy
   • Output columns, column order, column aliases
   • Boundary conditions, strict vs non-strict inequalities
   • Business-term-to-field mapping (e.g. which expression encodes "active" / "retained")
4. Treat each diff item as a candidate fact and pass it into Step 5's "worth-writing" filter.

There is no match / mismatch verdict in lite mode — diff items drive extraction directly. Lite mode does not enter the "5-round exhaustion" branch and does not write Open Gaps (no iteration history exists).

Step 2 (deep mode) — First subagent call

Take this branch only when mode = deep and continue with Steps 3 and 4 below.

Invoke:

task(
  type="gen_sql",
  prompt=<question>,         # natural-language question only
  description="extract-knowledge: initial attempt for <short topic>"
)

The returned envelope contains result.sql (or result.sql_file_path for long SQL), result.response, and result.session_id. Save the session_id — every later retry must reuse it.

Step 3 (deep mode) — Execute and compare

Run the subagent's SQL via execute_sql and compare against the cached gold result:

• Row count match?
• Column names / count match at the semantic level (aliases may differ)?
• After sorting both sides by the same key, do sampled rows match?
• For precise verdicts, use execute_sql to run difference probes (two-way EXCEPT, key-column SUM / COUNT(DISTINCT ...) consistency checks, etc.)

Verdict: match or mismatch.

Step 4 (deep mode) — Mismatch: diagnose and retry on the same session

Diagnose qualitatively. Common symptoms → likely missing knowledge:

When designing the follow-up prompt, obey:

• Describe symptoms and direction only. Never leak gold values. Never restate gold SQL's exact wording to the subagent.
• The "direction" portion of your prompt should already read like a draft of the future knowledge fact. I.e. when you write the prompt you have mentally passed it through the "worth-writing" filter: every business rule / mandatory filter / field trap you point the subagent at must be transcribable as a knowledge fact in Step 5 by simply switching from imperative to declarative voice. Benefits: prompts are shorter and more focused; Step 5 becomes near-zero-cost (cherry-pick validated directions from prompt history, strip imperative voice, write to file); prompts and knowledge content stay aligned.
• Avoid shotgun guessing ("maybe the join is wrong? or the aggregation? or null handling?") — such prompts cannot be transcribed into any fact and waste tokens.

Then call:

task(
  type="gen_sql",
  session_id=<saved session_id>,    # must be the previous round's id
  prompt=<hint only>,
  description="extract-knowledge: refine #<n>"
)

Loop Step 3 → Step 4 for at most 5 rounds (including the first). If still mismatched after round 5: stop retrying, but do not discard the process information — promote every confirmed knowledge gap (the rules the subagent finally got right each round, the business rules / field traps / mandatory filters you've already pinpointed in follow-up prompts) into facts via Step 5, then log the remaining un-aligned differences in Open Gaps via Step 7.

"Worth-Writing" Test (read before Step 5)

For every candidate diff item, ask these 4 questions in order. If any answer is "yes," drop it.

1. Without this knowledge, could a SQL agent with the question + schema still get it right?
2. Can this information be inferred directly from INFORMATION_SCHEMA / table comments / column names?
3. Is this generic SQL knowledge (not specific to this business / dataset)?
4. Can this fact be mechanically composed from other facts already in the file? (If two existing facts combine to produce it, the combination is derivative — do not write it.)

Record atomic facts only — anything else counts as derivative and is not written. Common atomic categories worth writing:

• Field encoding — bit-position meanings, enum mappings, status codes, special-value semantics
• Business measure definitions — the criteria behind business terms
• Mandatory constant filters — constant predicates whose omission corrupts results
• Boundary traps — strict inequalities, interval endpoint open/closed conventions
• Implicit table joins — relationships that must go through a mapping table; direct joins forbidden
• Same-name field divergence — the same field name has different semantics in different contexts
• Required timing / parameter constraints — business-mandated fixed parameters

Self-check: after writing a group of facts, look back — can you drop any single one and still let the LLM answer correctly? If yes, that one was derivative. Drop it.

Step 5 — Mine atomic facts (both modes)

Trigger:

• lite mode — enter immediately after Steps 2–4 (lite) produce the draft-vs-gold_sql diff.
• deep mode — match achieved or 5 rounds exhausted. Both cases enter — partial-alignment progress still carries knowledge value.

lite mode: every diff item from Steps 2–4 (lite) is a candidate fact. deep mode: diff the subagent's final SQL against gold_sql (also against the subagent's first attempt — the middle is informative). On failure, additionally walk through every follow-up prompt you sent; the "directions you've already qualitatively named" are confirmed fact sources.

Systematic corpus scan (both modes, additive) — mine encodings the per-pair diff misses. The diff only surfaces facts a generic agent got wrong on this question; a literal that the draft happened to guess, or that no pair exercised, slips through. So, when you have access to the validated-SQL corpus (not just the single pair), also scan its WHERE / SELECT clauses structurally for non-inferable atoms, independent of any single diff:

• Literal filter values / coded enumerations — every constant predicate (col = '<literal>', col IN (...), bit/flag tests) where the literal carries business meaning → a term ↔ column ↔ value fact (e.g. a business term maps to a specific column holding a specific code).
• Business-term → column / expression mappings — when a question's term resolves to a particular column or computed expression in SELECT/WHERE → record the mapping.
• Question-word → column trigger mappings — when a recurring interrogative phrasing (not just a noun) consistently resolves to one column or expression across pairs (e.g. one wording always hits column A while a near-synonym wording hits column B), record the trigger fact: question says ⟨phrasing⟩ → use ⟨column/expression⟩. These disambiguate near-duplicate columns that schema inspection cannot.
• Recurring mandatory constant filters — a constant predicate that appears across many pairs and whose omission corrupts results → a required-filter fact.
• Threshold facts always carry the comparison operator's strictness. A threshold without its boundary semantics is half a fact: record > 500 vs >= 500 (and BETWEEN's inclusivity) exactly as the validated SQL has it — strict-vs-inclusive mismatches flip rows at the boundary and are among the most common silent errors.

This pass extracts structure (literals in WHERE, term→column in SELECT), never hardcoded domain vocabulary — it works for any dataset. Every candidate it produces is just another candidate: funnel it through the same "worth-writing" test (so generic/inferable facts are still dropped) and the same Step 6 dedup/conflict gates (so it never duplicates what the diff already wrote).

Pass every candidate through the "worth-writing" test. Only survivors become facts.

Each fact contains only:

• statement — one sentence stating the fact itself. Do not explain "why it matters." Critical warnings (boundary values, easy-to-miss traps) go in inline parenthetical notes.
• example (optional) — only when the one-sentence statement cannot remove ambiguity by itself; then attach a minimal SQL snippet.

Forbidden:

• Restating the rule in SQL (don't say "use > 7, not >= 7" and then paste a WHERE x > 7 snippet)
• Writing "violating this would be wrong" as a standalone field — pure filler
• Writing as a standalone fact anything that mechanically composes from other facts

Knowledge File Layout

business domain   →   topic   →   facts
   (one .md)         (## heading)   (list items / table rows / paragraphs / optional ### subblock)

• Business domain = one file — a slice of business broad enough that its rules co-evolve. Path: ./knowledge/<domain-slug>.md. Prefer fewer, wider domains over many narrow ones.
• Topic (##) = a group of facts sharing context — typically the constraints of one table, a set of measure definitions, or a related mapping group.
• Fact representation, in priority order:
  1. Bulleted list item — multiple independent facts under the same topic (most common)
  2. Table row — several parallel field / enum / branch mappings
  3. Paragraph — a conceptual fact a sentence or two can explain
  4. ### <one-sentence rule> subblock — only when a SQL example must be attached

Hard constraints:

• A topic block must hold ≥2 independent facts (otherwise merge into a wider topic)
• Heading depth never exceeds ###
• Do not write Derived from / Why it matters / When it applies label fields
• The file must contain no question text (trace via git history instead)
• Strong preference: reuse first, create new last. Before persisting anything, list ./knowledge/ and read each file's Domain intro. Prefer adding a new topic or fact under an existing domain over opening a new file.

Template for a new business-domain file:

# <Domain Title>

> **Domain:** <one-sentence scope statement — what this file covers, what it doesn't>.

## <Topic Title>

<optional 1–2 sentence intro describing what this topic covers>

- fact 1
- fact 2 (note ...)

| Field / Branch | Value / Verdict |
|---------------|-----------------|
| ... | ... |

### <rule that needs a SQL example>

<one-sentence statement, may carry an inline warning>

\`\`\`sql
<minimal snippet that resolves ambiguity>
\`\`\`

Step 6 — Persist (fact-level dedup / conflict gates)

For every classified fact, walk the gates in order. Do not skip steps.

6.1 Resolve target domain file

• If ./knowledge/ does not yet exist, just write_file the first file (no ask_user needed for an empty-directory case).
• If the fact does not belong to any existing domain, before creating a new file you must ask_user: list existing domains, propose a new slug, let the user choose reuse / new / custom.
• Never silently open a new file when a reasonable reuse path exists.

6.2 Detect duplicates and conflicts inside the file

read_file the target. Extract existing facts per topic (list items + table rows + paragraph facts + ### subblock headings). Compare each candidate against existing facts semantically (not by string):

6.3 Write or edit the file

• Prefer edit_file for minimum diff.
• When a new fact changes the appropriate representation of an existing topic, whole-block rewriting that topic is allowed (only touch that block); avoid letting persistent append accumulate redundancy.
• Heading conventions:
  • New domain → write_file the Domain intro followed by the first topic heading
  • New top-level topic → insert a ## <topic title> block
  • New fact into existing topic → append as list item / table row / paragraph; only add a ### <one-sentence rule> subblock when a SQL example is needed

Do not insert --- separators between facts — markdown headings / lists already separate them, and --- makes edit_file insertion points ambiguous.

Step 7 — Update the AGENTS.md index

Maintain the ## Knowledge section of ./AGENTS.md, so that later main agents handling related business-domain tasks can enter from AGENTS.md, see what knowledge/ holds, and read_file the relevant domain file as context.

AGENTS.md handling order:

1. ./AGENTS.md missing → write_file a minimal skeleton (just # <project directory name> + the ## Knowledge section; leave the rest for /init to fill in later). Do not wait for the user to run /init first.
2. AGENTS.md exists but lacks the ## Knowledge section → insert after ## Artifacts (or at end of file if Artifacts is also absent).
3. Rescan ./knowledge/ and rewrite the entire ## Knowledge section sorted alphabetically by domain title — guarantees the index stays consistent across runs.

Fixed structure of the ## Knowledge section:

## Knowledge

`./knowledge/` holds this project's atomic business facts, organized per business domain (maintained by `/extract-knowledge`).
Before handling a task that touches a domain, `read_file` the corresponding file for context.

- [<Domain Title A>](knowledge/<domain-slug-a>.md) — <one-sentence scope>
- [<Domain Title B>](knowledge/<domain-slug-b>.md) — <one-sentence scope>

### Open Gaps

- <question summary> — <remaining un-aligned difference>

• One index line = one business-domain file (not per topic, not per fact). Scope sentence comes from that file's > Domain: intro.
• When ./knowledge/ is empty, omit the index list and ### Open Gaps; keep only the section intro paragraph.
• ### Open Gaps omits entirely when there are no entries; when present, group by domain.

5-round mismatch handling (deep mode only): keep the facts Step 5–6 already wrote in their domain files, then append a line - <question summary> — <remaining un-aligned difference> under ### Open Gaps, and continue with the next pair. Lite mode has no iteration history and does not write Open Gaps.

Final Output

Return a single human-readable summary covering: the mode used (lite / deep); how many pairs were processed; how many matched / failed (deep) or were aligned (lite); which knowledge/*.md files were created / edited; how many facts were added; any conflicts still needing attention. When lite was used because ask_user was unavailable, add one line: "defaulted to lite; rerun and choose deep for stricter validation." Do not return a JSON envelope.

Tools You'll Use

• execute_sql(sql=...) — execute gold SQL and difference probes (deep mode additionally executes subagent SQL). In lite mode it is only used for Step 1's gold validation.
• task(type="gen_sql", prompt=..., session_id=...) — deep mode only; delegate SQL generation, reuse session across retries.
• read_file, write_file, edit_file — manage ./knowledge/*.md and ./AGENTS.md. Always read_file the target domain file before edits (6.2 depends on it).
• ask_user — required for: ambiguous input parsing; confirming a context-recovered pair; mode selection (lite / deep); new-domain confirmation (6.1); refinement decisions (6.2); every fact conflict (6.2). Phrase questions with numbered options.

Forbidden

• In lite mode, do not call task(type="gen_sql", ...) — the mode selection already declared no subagent.
• In deep mode, do not call task(type="gen_sql", ...) before confirming gold_sql actually runs.
• In deep mode, do not open a new gen_sql session for retries — always pass the previous round's session_id.
• Do not write SQL on behalf of the user's question — that's the subagent's job (deep mode). Your role is orchestrator and knowledge curator.
• Do not put gold SQL into ./knowledge/*.md. Knowledge is atomic facts mined from the gap, not the answer itself.
• Do not write derivative facts — anything that mechanically composes from other facts is dropped.
• Do not write Derived from / Why it matters / When it applies label fields; do not include any question text in the files.
• Do not attach a SQL example to every fact — only when the one-sentence statement cannot remove ambiguity by itself.
• When an existing domain can plausibly cover the topic, do not create a new file. Reuse + slightly extending the intro is almost always right.
• Do not resolve fact conflicts silently. Conflicts must go through ask_user — the user decides replace / coexist / scope-gate.
• Do not insert --- separators between facts.
• Do not give rules generic names (### Rule 1, ### Note). The one-sentence rule itself is the heading.
• Topic heading depth must not exceed ####. Deeper means the domain was sliced wrong — fold the levels or split the file.
• A single-fact topic must not be its own ## — merge it into a wider topic.