> ## Documentation Index > Fetch the complete documentation index at: https://gladlabs.mintlify.site/llms.txt > Use this file to discover all available pages before exploring further. # Niche topic scope # Niche topic scope A niche can state what it covers. When it does, topic discovery keeps to that subject instead of taking whatever its sources happen to fetch. ([Glad-Labs/poindexter#1127](https://github.com/Glad-Labs/poindexter/issues/1127)) ## Why it exists The topic ranker scores candidates against a niche's **goals**: `AUTHORITY`, `EDUCATION`, `TRAFFIC` and so on. Goals say *why* to write something. Nothing said *what about*. The subject a niche ended up writing about was decided entirely by which sources fed its pool, so a niche meant to cover one field drifted into business, marketing and posts about itself whenever those came in. The only way to push back was editing `niche_goal_descriptions`, a single setting shared by every niche on the install. ## What you set Three columns on `niches`, set with `poindexter topics niche set-scope`: | Column | Meaning | | - | - | | `topic_subject` | What the niche covers, in plain prose. NULL means no scope: sweeps behave exactly as they did before scope existed. | | `topic_exclusions` | Subjects that are out of scope even when they sit next to the subject, such as "video game news" beside "PC hardware". | | `topic_scope_filter` | On (the default): out-of-scope candidates are dropped before ranking. Off: the subject only steers ranking. | ```bash theme={null} poindexter topics niche set-scope my-niche \ --subject "AI and computer hardware: models, training, inference, GPUs, CPUs, builds, servers" \ --exclude "video games and game news" \ --exclude "business strategy, marketing and startups" poindexter topics niche check-scope my-niche --show out # what would be dropped poindexter topics niche show my-niche # goals, sources, scope ``` `set-scope` edits only what its flags name, so adding one exclusion keeps the subject and the other exclusions. `--remove-exclude`, `--clear-exclusions`, `--filter/--no-filter` and `--clear` do what they say. ## What reads it * **The scope check (hard filter).** After the sanity filter and before embedding, the sweep sends each candidate's title and summary to `niche_topic_scope_check_model` (empty means the structured-extraction model) in batches of `niche_topic_scope_check_chunk_size` (15) and drops the ones whose main subject is out of scope. Carried-forward candidates are re-judged, so a scope edit applies to them on the next sweep. The prompt is `topic.scope_check` in `skills/content/research/SKILL.md`. * **The ranking prompt.** The scope is appended to the goal list the `topic.ranking` prompt receives, so the ranker scores off-subject candidates low even when the filter is off. It rides on the existing `weights_descr` variable rather than a new one, so a customised ranking prompt keeps working and still sees the scope. * **The `NICHE_DEPTH` goal.** With a subject set, `NICHE_DEPTH` is anchored on the subject text instead of the install-wide description, so it means "deep on this niche's subject". Weight it with `set-goals`. * **Internal story selection.** `internal_rag` passes the scope to its distiller and uses the same niche-aware goal vectors to pick snippets. ## Why an LLM check and not an embedding threshold Measured before building it, on 37 real titles against an AI + hardware subject with `nomic-embed-text`: title embeddings do not separate the two sets. "Market Research vs Industry Research" scored 0.478 against the subject, above "Decode Speed Lies: phi4:14b" at 0.477, and "Qwen 3.8 27B Needs 22,000 Tokens to Draw a Pelican" scored 0.450. Any threshold would drop real on-topic posts and keep off-topic ones. The same 37 titles in one batched call to qwen2.5:7b came back about 33 correct in 5.4 s. The embedding anchor is still used for `NICHE_DEPTH`, where it is one weighted signal among several rather than a gate. ## Which model, and how many per call Measured on 107 live candidates with 16 hand-labelled on-topic and 14 off-topic titles: | Model | Per call | On-topic wrongly dropped | Off-topic wrongly kept | | - | - | - | - | | qwen2.5:7b | 40 | 8 of 16 | 1 of 14 | | qwen2.5:7b | 15 | 4 of 16 | 1 of 14 | | qwen3-vl:30b-a3b-instruct | 15 | 0 and 1 of 16 (two runs) | 0 of 14 | Smaller models drop real on-topic candidates, often terse search-query titles such as "Gguf Quantization Types". Point `niche_topic_scope_check_model` at the strongest model you can afford to run on every sweep, and check the result with `check-scope` before relying on the filter. ## Failure behaviour The check **fails open and loud**. A model error, an unparseable reply or an id the model skipped keeps the candidate, and the sweep emits `topic_scope_check_failed` (routed to Discord, 6 h cooldown), so an off filter never looks like a working one. Each sweep that drops candidates emits one `topic_scope_filtered` finding listing them (logged and on the Findings board, not sent). If the filter drops everything, the existing empty-batch guard records the run and does not open an empty batch. `check-scope` prints the check's first error when anything is unjudged. A common one is running it from a host shell on an install whose model URL only resolves inside the containers (`host.docker.internal`). Run it in the worker instead: `docker exec poindexter-worker python -m poindexter topics niche check-scope `. ## Other niche controls ```bash theme={null} poindexter topics niche set-goals my-niche NICHE_DEPTH=35 AUTHORITY=25 EDUCATION=20 TRAFFIC=15 REVENUE=5 poindexter topics niche set-sources my-niche internal_rag=25 web_search=25 hackernews=20 devto=15 knowledge=5 poindexter topics niche set-writer-prompt my-niche --file writer-prompt.md ``` Goals must sum to 100. `set-sources` replaces the list; `NAME=WEIGHT:off` keeps a source listed but disabled. What each source fetches (queries, feeds, tags) lives on its `external_taps` row: `poindexter taps set-config`. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.