mmi-search

name

mmi-search

description

Search the MMI Elasticsearch `content` index, a BC government media-monitoring corpus of ~8.9M news clips (print, online, radio, TV, wire). Use this skill WHENEVER the user asks to find, list, count, or analyze media content, news stories, clips, or coverage in Elasticsearch, OR mentions any of these by name — top stories, featured stories, alerts, commentary, tone or sentiment, tags (ministries), topics, media types, or sources. Also use it for queries like "top stories today", "what coverage did X get", "negative sentiment about Y", "clips from the Vancouver Sun", or any filtering by ministry/tag, topic, date, source, or sentiment. This skill defines the schema and the correct (often non-obvious) query patterns; consult it before writing any query against the `content` index, because several fields require nested queries or `.keyword` subfields that fail silently if used wrong.

Searching the content index

The content index (MMI Elasticsearch connector) holds BC government media-monitoring records: individual news items from print, online, radio, TV, and wire sources, each enriched with editorial metadata (actions, tags, topics, sentiment).

Scale: ~8.9M documents. Always use size limits and date filters; never run an unbounded fetch.

Tooling

Use the MMI Elasticsearch:search tool with a Query DSL query_body. Use esql only for tabular aggregate reports. Get field types via MMI Elasticsearch:get_mappings (note: the full mapping is large and may fail to decode — prefer the schema in references/schema.md).

Critical: keep responses small

The connector fails to decode very large responses (you'll see error decoding response body). To avoid this:

• Always set _source to only the fields you need (or use the fields parameter).
• Keep size modest (≤100). Use aggregations with "size": 0 for counts/breakdowns.
• Never return the full _source of many documents at once.

Query gotchas (read before writing any query)

These are the mistakes that silently return 0 results or throw 400 errors:

1. actions, tags, topics, tonePools are nested objects. A flat {"match": {"actions.name": "Top Story"}} returns 0 hits. You MUST wrap them in a nested query, and to match a name+value pair on the same element, put both conditions inside one nested bool. See "Actions" below.
2. actions.value is text, not boolean. Match it with {"match": {"actions.value": "true"}}. term on actions.value.keyword returns 0.
3. Aggregations: some fields aggregate on the analyzed field, others need .keyword. Confirmed aggregatable: actions.name, tags.code, tags.name, topics.name, tonePools.name, tonePools.value, contentType, mediaTypeId, mediaType.name.keyword, source.name.keyword, status, licenseId. If an agg throws 400, try toggling .keyword.
4. Date boundaries are UTC. publishedOn is a UTC timestamp; BC print items often carry a 07:00:00Z stamp and TV/radio can spill into the next UTC day. For "today" (local), use a calendar range spanning the local day in UTC rather than now/d. See "Date filtering" below.
5. source.name sometimes has a leading space (e.g. " Victoria Times Colonist", " Kelowna Courier"). Prefer match over exact term for source name, or account for the space.

Schema

See references/schema.md for the full field list and types. The fields you'll use most:

Field	Type	Notes
id	long	Content id (also surfaced as externalUid/uid like tno-08949063)
headline	text	The headline (the property is singular: headline, not headlines)
summary	text/html	Short summary, may contain HTML
body	text	Full text (often empty for clips)
byline	text	Author
publishedOn	date	UTC publish timestamp — primary sort/filter field
postedOn	date	When ingested
contentType	keyword	PrintContent, Internet, AudioVideo, Image
mediaType.name	obj	e.g. Daily Print, Online, News Radio, TV / Video News (agg on mediaType.name.keyword)
mediaTypeId	int	numeric media type
source.name	text	Outlet name (agg on source.name.keyword; watch leading spaces)
source.code	—	Short source code
otherSource	keyword	Short outlet label, e.g. SUN, GLOBE, POST, TC, TS
series.name	obj	Program/series, e.g. CKNW, Global BC News Hour
status	keyword	Draft, Published, Unpublished
actions	nested	Editorial flags (Top Story, etc.) — see below
tags	nested	Ministry/subject tags — see below
topics	nested	Editorial topics — see below
tonePools	nested	Sentiment — see below
licenseId	int	Licensing bucket

Actions (editorial flags) — NESTED

Each element has name, value, valueType, valueLabel, defaultValue. Distinct action names and their value types:

Action name	valueType	Meaning
Featured Story	Boolean	Flagged as a featured story
Top Story	Boolean	Flagged as a top story
Alert	Boolean	Flagged for alerting
Non Qualified Subject	Boolean	Editorial qualifier flag
Commentary	String	Commentary marker (value is a string label, e.g. timeout)

Boolean actions are "on" when value == "true" (string). To find documents where a given action is on, match name and value within the same nested element:

{
  "query": {
    "nested": {
      "path": "actions",
      "query": {
        "bool": {
          "must": [
            { "match": { "actions.name": "Top Story" } },
            { "match": { "actions.value": "true" } }
          ]
        }
      }
    }
  }
}

Tags (ministries / subjects) — NESTED

Each element has code, name, id, contentId. Tags map to BC government ministries and subject areas. Examples (code → name):

PSSG → Public Safety & Solicitor General · HLTH → Health · MJAG → Attorney General · ENV → Environment and Parks · TRAN → Transportation and Transit · JTST → Jobs, Economic Development & Innovation · FIN → Finance · FORR → Forests · EMBC → Emergency Management · EDU → Education & Child Care · TACZ → Tourism, Arts, Culture & Sport · MSD → Social Development & Poverty Reduction · TNF → Indigenous Relations & Reconciliation · HOZZ → Housing and Municipal Affairs · PDPE → Premier David Eby · LBRR → Labour · AGG → Agriculture & Food · MCFD → Children & Family Development · MHAA → Mental Health & Addictions · ECS → Energy and Climate Solutions · MCMX → Mining and Critical Minerals · ZHJ → South Asian Media · ZHX → Chinese language media · XRZ → Transcripts.

TNO is by far the most common code; many older records have the placeholder name -- UPDATE MIGRATED TAG NAME --, so filter/aggregate on tags.code for reliability. Full list in references/schema.md.

{ "query": { "nested": { "path": "tags", "query": { "term": { "tags.code": "HLTH" } } } } }

Topics — NESTED

Each element has name (and id). Topics are editorial story-threads, e.g. Wildfire season, Trade war, Overdose crisis, Housing shortage, FIFA 2026, Cost of living, Toxic drug crisis. Not Applicable is the most common. Many COVID-era topics exist (COVID-19 vaccine, etc.).

{
  "query": {
    "nested": { "path": "topics", "query": { "match": { "topics.name": "Wildfire season" } } }
  }
}

Tone pools (SENTIMENT) — NESTED

tonePools encodes sentiment. Each element has name (almost always Default), value (the sentiment score), isPublic, ownerId. The scoring scale observed:

value	meaning
> 0 (1…5)	Positive (higher = more positive)
0	Neutral (most common)
< 0 (-1…-5)	Negative (lower = more negative)
-99	Sentinel / not scored (exclude from analysis)

To find negative-sentiment items (using the public Default pool):

{
  "query": {
    "nested": {
      "path": "tonePools",
      "query": {
        "bool": {
          "must": [
            { "term": { "tonePools.name": "Default" } },
            { "range": { "tonePools.value": { "lt": 0, "gt": -99 } } }
          ]
        }
      }
    }
  }
}

For a sentiment breakdown, aggregate tonePools.value inside a nested agg (remember to exclude -99).

Date filtering

publishedOn is UTC. The user's local zone is Pacific (Surrey, BC). For a local "today"/"yesterday", compute the UTC span explicitly rather than relying on now/d, because BC print items are stamped 07:00:00Z and broadcast items can land after UTC midnight. A practical pattern that captures a local day's coverage:

{ "range": { "publishedOn": { "gte": "2026-06-05T00:00:00Z", "lte": "2026-06-06T23:59:59Z" } } }

When the user just says "today", confirm the date, build the range around it, sort by publishedOn desc, and note the UTC caveat if results look off.

Recipe: Top stories for a day

Combine the nested Top-Story flag with a date range and a trimmed _source:

{
  "_source": ["id", "headline", "publishedOn", "otherSource", "source.name"],
  "query": {
    "bool": {
      "must": [
        {
          "nested": {
            "path": "actions",
            "query": {
              "bool": {
                "must": [
                  { "match": { "actions.name": "Top Story" } },
                  { "match": { "actions.value": "true" } }
                ]
              }
            }
          }
        },
        { "range": { "publishedOn": { "gte": "<DAY>T00:00:00Z", "lte": "<DAY+1>T23:59:59Z" } } }
      ]
    }
  },
  "size": 100,
  "sort": [{ "publishedOn": { "order": "desc" } }]
}

Present results as a table with id, headline, publishedOn, source (otherSource + source.name). More worked recipes (sentiment over time, coverage by ministry, source breakdowns, counting) are in references/recipes.md.

Output

Default to a compact table unless the user asks otherwise. Include the document id so items can be looked up. When you filtered on a nested field, briefly note it (e.g. "Top Story flag is a nested action") only if the user is debugging — otherwise just answer.