Signals

Querying Signals

What Are Signals?

Signals are automated observations that PostHog generates by monitoring a customer’s product data across multiple sources — error tracking, web analytics, experiments, session replay, and more. Each signal is a short natural-language description of something noteworthy (e.g. “Error rate spiked 3× on /checkout”).

Signals are grouped into Signal Reports. When a report accumulates enough weight it gets summarized and assessed for actionability. A signal report represents a cluster of related observations that together describe a meaningful issue or trend.

Signals and their embeddings are stored in the document_embeddings ClickHouse table, queryable via HogQL through the posthog:execute-sql MCP tool. They may provide a useful way to semantically query for recent things that happened in the user’s product.

When to use this skill vs. inbox-exploration

The two skills cover different layers of the same product:

• inbox-exploration — curated report layer via dedicated MCP tools (inbox-reports-list, inbox-reports-retrieve, inbox-source-configs-list, inbox-source-configs-retrieve). Use for “what’s in my inbox?”, “what’s actionable?”, filtering reports by status / source / suggested reviewer, looking up a specific report by ID or URL.
• This skill (signals) — raw signal layer via HogQL on document_embeddings. Use when the curated report layer is not enough: semantic search over signal text, fetching every signal that contributed to a specific report, listing what kinds of signals exist, or any ad-hoc analytics that the report tools don’t expose.

The typical pattern is to start with inbox-exploration, get a report_id or a sense of the area the user cares about, then drop into this skill when the user wants to see the raw observations.

Table and Column Reference

The HogQL table alias is document_embeddings. HogQL automatically constrains queries to the current team — you never need to filter on team_id. Key columns for signals:

Mandatory Filters

Every signals query MUST include all four of these filters. Missing any of them can cause the query to fail with an invalid model error, return wrong data, or trigger unnecessarily expensive scans:

WHERE model_name = 'text-embedding-3-small-1536'
  AND product = 'signals'
  AND document_type = 'signal'
  AND timestamp >= now() - INTERVAL 30 DAY

The model_name filter is especially critical — the HogQL engine uses it to route to the correct underlying ClickHouse table. If the WHERE model_name = ... equality filter is missing or uses an unknown model, the query will fail with an “Invalid model name” error (you cannot use IN or other expressions here).

The product and document_type filters are equally important — the same model contains data from multiple products (e.g. error tracking, AI memory). Without these filters you will get unrelated data mixed in.

The timestamp filter is required for performance — the table is partitioned by week and has a 3-month TTL. Always include a time bound using now() - INTERVAL N DAY (or WEEK, MONTH, etc.). Default to 30 days unless you have a reason to look further back. Generally, more recent data is more likely to be relevant, unless investigating a long-standing issue.

Deduplication Pattern

The underlying table can contain multiple versions of the same signal (e.g. after a soft-delete re-emission). You MUST always deduplicate by wrapping reads in a subquery using argMax(..., inserted_at) grouped by document_id.

Note: HogQL supports metadata.field_name dot access on the raw metadata JSON column, but this type information is lost when the column passes through aggregate functions like argMax(). You MUST extract individual metadata fields inside the inner dedup subquery — do NOT pass the whole metadata blob through argMax and dot into it in the outer query, as this will fail with a type error.

HogQL’s JSON dot access always extracts values as Nullable(String), regardless of the underlying JSON type. This means metadata.deleted is the string 'true'/'false'/null, not a Bool. Use deleted != 'true' — do NOT use NOT deleted.

SELECT ... FROM (
    SELECT
        document_id,
        argMax(content, inserted_at) as content,
        argMax(metadata.report_id, inserted_at) as report_id,
        argMax(metadata.source_product, inserted_at) as source_product,
        argMax(metadata.source_type

Only select the embedding column in the inner subquery when you actually need it for similarity searches — it’s a 1536-element float array and expensive to materialize otherwise.

The embedText() Function

embedText() is a HogQL function that converts a text string into an embedding vector at query compile time. It calls the embedding API and inlines the resulting vector as a constant before executing the query. This means you can do semantic search in a single query without any external embedding step.

Signature: embedText(text, model_name)

• text — the string to embed. Must be a string literal, not a column reference.
• model_name — the embedding model to use. For signals, always use 'text-embedding-3-small-1536'.

Both arguments must be literal strings. You cannot pass column values or expressions — the function resolves at compile time, not per row.

cosineDistance() for Similarity Search

Use cosineDistance(embedding, ...) to rank signals by semantic similarity. Lower values = more similar. Always ORDER BY distance ASC and add a LIMIT.

cosineDistance(embedding, embedText('your search text', 'text-embedding-3-small-1536')) as distance

The embedding model (text-embedding-3-small-1536) uses matryoshka representation learning, so the embedding dimensions are ordered by importance. This means similarity search works well even at high dimensionality — the curse of dimensionality is not a significant concern here.

Metadata JSON Fields

The metadata column is a JSON string. HogQL supports metadata.field_name dot access only on the raw table column. After aggregation (e.g. argMax), the JSON type is lost and dot access will fail. Always extract the fields you need inside the dedup subquery.

Example 1: Semantic Search for Signals

Find signals most similar to a natural-language query. This is the most useful query for understanding what’s happening in a customer’s product:

SELECT
    document_id,
    content,
    report_id,
    source_product,
    source_type,
    cosineDistance(embedding, embedText('users seeing errors on checkout page', 'text-embedding-3-small-1536')) as distance
FROM (
    SELECT
        document_id,
        argMax(content, inserted_at) as content,
        argMax(

Adjust the embedText first argument to whatever you’re looking for. Write it as a natural-language description of the kind of issue or observation you want to find.

To restrict to signals that have already been grouped into a report, add AND report_id != '' to the outer WHERE.

Example 2: Fetch All Signals for a Specific Report

Once you have a report_id (from a semantic search or from the Signal Reports API), fetch all signals belonging to that report:

SELECT
    document_id,
    content,
    report_id,
    source_product,
    source_type,
    signal_ts
FROM (
    SELECT
        document_id,
        argMax(content, inserted_at) as content,
        argMax(metadata.report_id, inserted_at) as report_id,
        argMax(

Example 3: List Signal Types

See what kinds of signals exist for this customer — returns one example per unique (source_product, source_type) pair from the last month:

SELECT
    source_product,
    source_type,
    count() as cnt,
    max(signal_ts) as latest_timestamp
FROM (
    SELECT
        document_id,
        argMax(metadata.source_product, inserted_at) as source_product,
        argMax(metadata.

Example 4: Recent Signals from a Specific Source

Find the latest signals from a particular product source (e.g. all error tracking signals):

SELECT
    document_id,
    content,
    source_type,
    report_id,
    signal_ts
FROM (
    SELECT
        document_id,
        argMax(content, inserted_at) as content,
        argMax(metadata.source_product, inserted_at) as source_product,
        argMax(metadata.

Replace 'error_tracking' with any source product: 'web_analytics', 'experiments', 'session_replay', etc. Use Example 3 to discover what source products and types exist.

Example 5: Full-Text Search for Signals

When you know a specific keyword or phrase to search for (e.g. a product name, error message, or URL), full-text search with ILIKE is faster and more precise than semantic search:

SELECT
    document_id,
    content,
    source_product,
    source_type,
    signal_ts
FROM (
    SELECT
        document_id,
        argMax(content, inserted_at) as content,
        argMax(metadata.source_product, inserted_at) as source_product,
        argMax(metadata.source_type

Replace '%feature flag%' with whatever term you’re looking for. Use ILIKE for case-insensitive substring matching. For exact token matching, use hasTokenCaseInsensitive(content, 'token') instead.

Gotchas

1. Always use text-embedding-3-small-1536 as the model name. This is the only model used for signals.
2. embedText() arguments must be string literals. You cannot pass column references or expressions — the function resolves at compile time, not per row.
3. Always time-bound your queries. The table has a 3-month TTL, but unbounded scans are expensive. Use timestamp >= now() - INTERVAL 1 MONTH or tighter. Place the time filter in the inner subquery’s WHERE clause (on the raw timestamp column) for best performance.
4. Always deduplicate. Without the argMax(..., inserted_at) GROUP BY document_id subquery, you will see stale and duplicate rows.
5. Only select embedding when you need it. It’s a 1536-element float array — omit it from the inner subquery when you’re not doing similarity search.
6. Queries should not end with a semicolon. HogQL does not use them.
7. Add a LIMIT to every query. Maximum allowed is 500 rows. In general, you should only select 10 or so signals, using semantic or full text search to rank them.
8. Extract metadata fields inside the dedup subquery. HogQL’s metadata.field dot access only works on the raw table column. After argMax() aggregation, the JSON type is lost and dot access will fail with a type error. Always use argMax(metadata.field_name, inserted_at) as field_name in the inner query.
9. All JSON dot-access values are Nullable(String). HogQL extracts every JSON field as a String, even booleans and numbers. For metadata.deleted, use deleted != 'true' — do NOT use NOT deleted.
10. Don’t alias argMax(timestamp, inserted_at) as timestamp if the same inner query also filters on the raw timestamp column. HogQL resolves the alias name first, causing an “aggregate in WHERE” error. Either use a distinct alias like signal_ts, or move the time filter to the outer query.