agentsclimarketplace

Index query engine

Skill cardstack/boxel/.claude/skills/index-query-engine

Architecture of the index query engine (`packages/runtime-common/index-query-engine.ts`) — the two-pass pipeline that compiles a `Query` filter tree into SQL that runs identically on Postgres and SQLite, plus the invariants every change must preserve (deferred expression nodes, cardinality-driven json_tree confinement, filter polarity, the shared table-valued alias semantics, synthetic search-doc keys, and the client-side parity contract with `instance-filter-matcher.ts`). Use when editing or reviewing the query engine, adding a filter operator or synthetic key, changing how field paths compile, or debugging why a query behaves differently across adapters or between server and local search.From its SKILL.md

Install
npx -y skills add cardstack/boxel --skill index-query-engine

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

SKILL.md

13.9 KB, as published. Nobody here has run it

Index query engine — architecture

The engine (packages/runtime-common/index-query-engine.ts) ingests a Query (filter tree + sort + page, defined in query.ts) and compiles it to one SQL statement that must run identically on Postgres and SQLite. Everything about its design follows from two constraints:

  1. Field semantics live in card definitions, which resolve asynchronously (DefinitionLookup — a cache read, or a module prerender on a miss). SQL can't be produced until the schema walk happens.
  2. SQLite is the floor. No arrays, no unnest, no @> containment, no ANY_VALUE, single-expression COUNT(DISTINCT …). Plural (containsMany / linksToMany) traversal therefore uses json_tree() — the lowest common denominator — and Postgres mirrors it with a custom jsonb_tree() function so one generated query shape serves both.

The pipeline

Query (filter tree)
  │  filterCondition()            stage 0 — sync tree expansion
  ▼
CardExpression                    tokens + DEFERRED schema-dependent nodes
  │  makeExpression()             pass 1 — async, definition-driven
  ▼
Expression                        only strings, params, resolved nodes
  │  expressionToSql()            pass 2 — mechanical rendering (expression.ts)
  ▼
SQL text + binds  ──sqlite──▶  SqliteAdapter.adjustSQL() text rewrites

Stage 0 (filterCondition and the per-operator builders) converts the filter tree into a CardExpression: a flat token array of SQL strings, param() bind markers, and deferred nodes. This is where a predicate is separated into left side / operator / right side:

  • left = fieldQuery(path, onRef, …) — the dotted field path, to become a JSON accessor;
  • operator = a literal (=, IS NULL, ILIKE, IN (…), range operators);
  • right = fieldValue(path, [param(v)], onRef, …) — deferred so the leaf field's serializer can format the bind (formatQuery);
  • the whole predicate wraps in fieldArity({ type, path, value, … }), which defers the plural-vs-singular decision.

Pass 1 (makeExpression) resolves the deferred node kinds against definitions. Each handler walks the dotted path segment-by-segment with walkFilterFieldPath, resolving a FieldDefinition per hop (getFieldDef chases each segment's fieldOrCard code ref) and tagging plural segments with [] in the traveled path:

  • handleFieldArity — if the walk crossed a plural segment, ANDs the predicate with the query-path anchor that confines json_tree()'s recursive expansion to exactly the intended path. This is the engine's most nuanced mechanism — see the dedicated section below.
  • handleFieldQuery — builds the left accessor. The forward (enter) walk detects the first plural hop and switches the whole left side to the tree alias; only if no plural was hit does the backward (exit) walk assemble search_doc -> 'a' ->> 'leaf'. Numeric leaves get a pg-only ::numeric cast (SQLite's ->> preserves JSON types).
  • handleFieldValue — runs the bind value through the leaf field's serializer so it compares in indexed form.
  • handleJsonContainsQuery — promotes a positive-polarity singular string eq to a JsonContains node (pg: GIN-servable search_doc @> '{…}'::jsonb; sqlite: plain ->/->> extraction); other leaves degrade to ->> equality.

Pass 1 also enforces searchability: filter paths (never sort paths) are walked with the queried card's searchable routes, and crossing an unannotated linksTo/linksToMany (except a bare .id hop) or a query-backed relationship raises FilterRefersToNonsearchableFieldError instead of silently matching nothing.

Pass 2 (expressionToSql in expression.ts) is purely mechanical: params become $n binds; table-valued nodes dedupe into a map (each_<column>, tree_<column>_<fieldPath>) rendered as nonce-numbered aliases, with the CROSS JOIN LATERAL … clauses spliced into the FROM clause at the __TABLE_VALUED_FUNCTIONS__ placeholder; dbExpression nodes pick their pg/sqlite branch.

Plural paths: json_tree and the query-path anchor

->/->> navigation cannot cross an array of unknown length, so any path that traverses a plural field routes through a json_tree table function. SQLite's is built-in; Postgres's jsonb_tree(data, root_path) is a custom recursive-CTE function (defined in the initial migration in packages/postgres/migrations/) returning (fullkey, jsonb_value, text_value, level) — deliberately emitting the same fullkey dialect as SQLite ($.friends[0].bestFriend.name: array hops bracketed, object hops dotted). That shared dialect is what lets one anchor pattern serve both adapters.

Worked example — friends is linksToMany, filter { eq: { 'friends.bestFriend.name': 'Mango' } }:

  1. The pass-1 walk tags plural segments, yielding the traveled path friends[].bestFriend.name, and roots the tree at the smallest prefix containing the fan-out: trimPathAtFirstPluralField → $.friends.
  2. json_tree enumerates every node beneath that root — one row each for $.friends, $.friends[0], $.friends[0].name, $.friends[0].bestFriend, $.friends[0].bestFriend.name, $.friends[1]… A value predicate alone (tree.text_value = $1) would therefore match 'Mango' appearing anywhere under $.friends — a friend's own name, a nickname field, any depth.
  3. handleFieldArity ANDs the anchor: tree.fullkey LIKE '$.friends[%].bestFriend.name' — the traveled path with each [] rewritten to [%] (convertBracketsToWildCards), so any array index passes but the key sequence must match exactly. The conjunction reads: "∃ a tree row at exactly this path shape whose value matches".
SELECT url, ANY_VALUE(pristine_doc) AS pristine_doc
FROM boxel_index AS i
CROSS JOIN LATERAL jsonb_tree(i.search_doc, '$.friends') AS friends0_tree
WHERE (i.is_deleted = FALSE OR i.is_deleted IS NULL)
  AND ( friends0_tree.text_value = $1                            -- value
        AND friends0_tree.fullkey LIKE '$.friends[%].bestFriend.name' ) -- anchor
GROUP BY i.url

One alias, two references. The left side (handleFieldQuery emits tree.text_value) and the anchor (handleFieldArity emits tree.fullkey) both create tableValuedTree(column, rootPath, fieldPath, …) nodes, and pass 2 dedupes them by tree_<column>_<fieldPath> — so they collapse into a single CROSS JOIN LATERAL and are guaranteed to test the same tree row. This is why handleFieldQuery computes its rootPluralPath to line up with handleFieldArity's: if the two emitted different (column, fieldPath) pairs they would get separate joins, and the anchor would no longer confine the value test.

Consequences of that dedup key (the full dotted field path):

  • Two predicates on the same path share one alias and evaluate per tree row: every of friends.name = 'Mango' and friends.name = 'Ringo' matches nothing, even for a card that has both friends — the same exists-one-element behavior as the shared types alias (next section).
  • Predicates on different leaf paths get separate tree functions and fan out as a cartesian product: friends.name = 'Mango' AND friends.age = 5 means "some friend named Mango AND some friend aged 5" — not necessarily the same friend. Element-correlated predicates across a plural hop are not expressible today.

Null checks anchor at the container (usePluralContainer). For eq/in/contains null on a path whose leaf is plural, there may be no element rows at all, so the anchor is redirected at the array node itself: trimTrailingBrackets drops the trailing [] (interior plurals keep their [%]), and the tree's seed row — jsonb_tree emits the root node as its first row — is what the anchor then matches. The comparison switches from text extraction to pluralValue: tree.jsonb_value = 'null'::jsonb (JSON null; adjustSQL rewrites this to tree.value IS NULL for SQLite). This is also what FieldQuery.useJsonBValue selects: jsonb_value for these JSON-typed comparisons, text_value for string predicates (SQLite maps both to json_tree's single .value column).

The two-adapter split

Structural differences are handled in-band (a dbExpression branch, a per-kind param, or a resolved node rendered per adapter — JsonContains is the precedent to copy when a predicate needs genuinely different SQL per adapter). Surface-syntax differences are patched by text rewrites in SqliteAdapter.adjustSQL (packages/host/app/lib/sqlite-adapter.ts): jsonb_tree(→json_tree(, jsonb_array_elements_text(→json_each(, ANY_VALUE( stripped via balanced-paren scan, CROSS JOIN LATERAL→CROSS JOIN, ILIKE→LIKE, .text_value/.jsonb_value→.value, = 'null'::jsonb→IS NULL, COLLATE "POSIX" dropped. If you emit new SQL through the shared path, check whether adjustSQL needs to learn it — and prefer a resolved node over a new text rewrite when the shapes diverge.

Semantics to keep in mind

Fan-out + GROUP BY = existential predicates. Table-valued functions fan each index row into one row per JSON element; WHERE filters those; GROUP BY i.url keeps any url with ≥1 surviving row. So every predicate that references a table-valued alias means "some element satisfies this".

All references to the same column share ONE alias (pass 2 dedupes by column/path). Consequences for types conditions specifically:

  • every([{type: A}, {type: B}]) requires a single element equal to both — unsatisfiable for distinct types, even though a card's types chain contains both. Positive type conditions do not AND-compose.
  • not: {type: X} means "some element ≠ X", which nearly every row satisfies — it does not exclude instances of X.
  • A row whose types is empty or null vanishes from any query that references the alias in any branch (the cross join produces zero rows for it), even under OR/NOT. Errored instances can lack types.

Don't build features that depend on type-condition composition without first changing typeCondition to a self-contained membership predicate (pg types @> '["key"]'::jsonb, sqlite EXISTS (SELECT 1 FROM json_each(types) WHERE value = ?) — a resolved node, per the JsonContains pattern).

Filter polarity. not flips the tracked polarity, and the @> containment rewrite for eq applies only at positive polarity: on an absent path ->> yields SQL NULL while @> yields FALSE, and NOT NULL vs NOT FALSE diverge. Any new operator that offers an index-servable rewrite must make the same polarity distinction.

Synthetic search-doc keys. Some queryable keys are not card fields — they are stamped into search_doc at index time (e.g. _cardType, _title) and resolved by the shim in this module's getField fallback. eq: null on such a key compiles to IS NULL, which matches rows that lack the key — mind that when a key is stamped on only some row kinds.

The HTML channel. Reads LEFT JOIN prerendered_html (ph) 1:1 on the primary key; ph is the sole home of rendered output — the HTML formats, markdown (the FTS matches predicate reads ph.markdown), the deps carrying scoped-CSS URLs, and the rendering generation. A row with no ph row reads them as NULL: no rendering exists yet, so it has no full-text membership either. icon_html stays on boxel_index (the icon renders in the index visit). The effective error state spans both channels (effectiveHasError). The join never fans out the GROUP BY.

The client-parity contract

packages/runtime-common/instance-filter-matcher.ts is this engine's client-side mirror: the host's live search (SearchResource.displayedInstances in packages/host/app/resources/search.ts) evaluates the same Filter/Sort against hydrated instances for immediate results, then reconciles the server response — and it removes a server result that evaluates locally to no-match (only unresolvable is trusted to the server). Therefore:

  • A new filter operator must be added to the matcher, or excluded in its isClientEvaluable so queries using it fall back to server-only evaluation.
  • A new synthetic key must get a matcher-side shim producing the value the index stamps (or evaluate as unresolvable), or every local instance returns no-match and reconciliation blanks correct server results.
  • Serializer/formatQuery changes must stay mirrored — the matcher formats filter values through the same card-api serializers on purpose.

Treat index-query-engine.ts and instance-filter-matcher.ts as one contract: a semantic change to either ships with its counterpart in the same change.

Key files

  • packages/runtime-common/index-query-engine.ts — stage 0 + pass 1
  • packages/runtime-common/expression.ts — node kinds + pass 2
  • packages/runtime-common/query.ts — Query/Filter types + assertions
  • packages/runtime-common/definitions.ts — FieldDefinition, getFieldDef
  • packages/host/app/lib/sqlite-adapter.ts — adjustSQL text rewrites
  • packages/runtime-common/instance-filter-matcher.ts — client-side mirror
  • packages/host/tests/unit/index-query-engine-test.ts — engine behavior tests

Keep looking

Skills are one crate of 613,757. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.