tdsql-mcp 1.3.2__tar.gz → 1.3.4__tar.gz

{tdsql_mcp-1.3.2 → tdsql_mcp-1.3.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tdsql-mcp
-Version: 1.3.2
+Version: 1.3.4
 Summary: MCP server for Teradata Vantage — SQL execution and native analytics function reference for AI agents
 Project-URL: Homepage, https://github.com/ksturgeon-td/tdsql-mcp
 Project-URL: Repository, https://github.com/ksturgeon-td/tdsql-mcp

{tdsql_mcp-1.3.2 → tdsql_mcp-1.3.4}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "tdsql-mcp"
-version = "1.3.2"
+version = "1.3.4"
 description = "MCP server for Teradata Vantage — SQL execution and native analytics function reference for AI agents"
 readme = "README.md"
 requires-python = ">=3.10"

{tdsql_mcp-1.3.2 → tdsql_mcp-1.3.4}/skills/teradata-sql-analytics/syntax/guidelines.md

@@ -195,6 +195,31 @@ Native functions distribute across all AMPs. The result set returned to the agen
 | External TF-IDF | `TD_TFIDF` | `text-analytics` |
 | External word embeddings | `TD_WordEmbeddings` | `text-analytics` |
 
+### LLM-Powered Text Analytics (AI_* Functions)
+
+> **Prerequisites:** Requires an authorization object and LLM provider configuration before use. See `authorization-objects` and `llm-providers` topics. All functions use `TD_SYSFNLIB.<FunctionName>(ON ...)` — do **not** add a `PARTITION BY` clause.
+
+| Instead of this | Use this (native function) | Topic |
+|-----------------|---------------------------|-------|
+| External/post-hoc sentiment scoring | `AI_AnalyzeSentiment` | `ai-text-analytics` |
+| Per-row LLM API calls with question + context data | `AI_AskLLM` (two-table: InputTable + ContextTable, co-partitioned by key) | `ai-text-analytics` |
+| External language detection | `AI_DetectLanguage` | `ai-text-analytics` |
+| External key phrase extraction | `AI_ExtractKeyPhrases` | `ai-text-analytics` |
+| External PII masking | `AI_MaskPII` (detects PII + returns `Masked_Phrase` with `*` replacement) | `ai-text-analytics` |
+| External NER (general named entities — people, places, orgs, dates) | `AI_RecognizeEntities` | `ai-text-analytics` |
+| External PII entity detection (structured metadata, no masking) | `AI_RecognizePIIEntities` | `ai-text-analytics` |
+| External text classification with custom label set | `AI_TextClassifier` — supports single-label and multi-label | `ai-text-analytics` |
+| External summarization | `AI_TextSummarize` — supports 1–5 compression levels | `ai-text-analytics` |
+| External translation | `AI_TextTranslate` | `ai-text-analytics` |
+
+**NER/PII function selection:**
+
+| Need | Function |
+|------|----------|
+| General named entities (people, places, orgs, dates) | `AI_RecognizeEntities` |
+| PII detection + masked text output | `AI_MaskPII` |
+| PII detection + structured metadata only (no masking) | `AI_RecognizePIIEntities` |
+
 ### Vector Search
 
 | Instead of this | Use this (native function) | Topic |
@@ -211,6 +236,18 @@ Native functions distribute across all AMPs. The result set returned to the agen
 
 **Finding the embedding model for an existing corpus:** Query `TD_SYSAI.TD_CollectionsV` or `TD_SYSAI.TD_VectorStores` to discover the model name, provider, and embedding size used to build a corpus. The query pipeline must use the exact same model — mismatched embeddings produce meaningless scores. See `vector-search` topic, "Discovering Existing Vector Stores" section.
 
+### Embeddings
+
+| Scenario | Use this | Topic |
+|----------|---------|-------|
+| REST-based embedding API (Azure, AWS Bedrock, GCP, NVIDIA NIM, LiteLLM) | `AI_TextEmbeddings` with `OutputFormat('VECTOR')` | `embeddings` |
+| In-database inference — no external API (air-gapped or latency-sensitive) | `ONNXEmbeddings` — model stored as BLOB in Vantage; requires tokenizer table | `embeddings`, `byom-model-loading` |
+| Classical word/document embeddings (GloVe-style) | `TD_WordEmbeddings` | `text-analytics` |
+| Store embeddings for reuse | CTAS with `VECTOR` column, then `TD_VectorNormalize(Approach('UNITVECTOR'))` at storage time | `embeddings`, `vector-search` |
+| Build fast approximate search index | `TD_HNSW` on normalized VECTOR column | `vector-search` |
+
+> **Always use `OutputFormat('VECTOR')`** when embeddings will be stored, normalized, or used with `TD_VectorDistance` / `TD_HNSW` / `TD_HNSWPredict`. The `VECTOR` type integrates directly with all vector search functions. See `data-types-casting` for VECTOR sizing (bytes, not dimensions).
+
 ### JSON Data
 
 | Operation | Use this | Topic |
@@ -236,6 +273,23 @@ Native functions distribute across all AMPs. The result set returned to the agen
 | Vantage ARRAY → JSON | `ARRAY_TO_JSON(arr_col)` | `json-functions` |
 | ST_Geometry ↔ GeoJSON | `GeoJSONFromGeom(geom)` / `GeomFromGeoJSON(json, srid)` | `json-functions` |
 
+### BYOM — Bring Your Own Model
+
+Apply externally trained models to in-database data without moving data out of Teradata. All scoring functions share the same two-table pattern: `InputTable` + `ModelTable DIMENSION`. See `byom-model-loading` topic for model ingestion; see `byom-scoring` for full syntax.
+
+| Instead of this | Use this (native function) | Topic |
+|-----------------|---------------------------|-------|
+| Running PMML model inference externally | `PMMLPredict` | `byom-scoring` |
+| Running H2O MOJO or Driverless AI model externally | `H2OPredict` (supports contributions, stage probabilities, leaf node assignments) | `byom-scoring` |
+| Running ONNX tabular model externally | `ONNXPredict` (use `ShowModelInputFieldsMap('true')` to inspect tensor mapping) | `byom-scoring` |
+| Running Dataiku Thin JAR externally | `DataikuPredict` (model_id = fully qualified Java class name) | `byom-scoring` |
+| Running DataRobot Scoring Code externally | `DataRobotPredict` (cast DATE/TIMESTAMP to VARCHAR before scoring) | `byom-scoring` |
+| Running MLeap model externally | `MLeapPredict` | `byom-scoring` |
+| In-database text generation / seq-to-seq (translation, summarization) | `ONNXSeq2Seq` — ONNX transformer, no external API | `byom-scoring` |
+| In-database text classification (transformer) | `ONNXClassification` — supports softmax, argmax, custom output column mapping | `byom-scoring` |
+
+> **Architecture distinction:** `ONNXSeq2Seq` and `ONNXClassification` run Hugging Face ONNX transformer models entirely in-database. For REST-based LLM text tasks (sentiment, PII, translation, summarization), use the `AI_*` functions in `ai-text-analytics` instead.
+
 ### Statistical Testing
 
 | Instead of this | Use this (native function) | Topic |

{tdsql_mcp-1.3.2 → tdsql_mcp-1.3.4}/skills/teradata-sql-analytics/syntax/sql-basics.md

@@ -140,15 +140,70 @@ CREATE USER myuser AS
 -- Semicolons: required in BTEQ; optional in most client tools
 ```
 
-## Reserved Words as Column Names in Table Operator Clauses
+## Reserved Words and Identifier Quoting
 
-When a column name is a Teradata reserved word (e.g. `type`, `date`, `time`, `value`, `name`, `format`, `title`), it must be double-quoted. In regular SQL projections this looks normal:
+Quote any identifier (column, table, alias) that conflicts with a reserved word using **double quotes**:
 
 ```sql
-SELECT "type", "date" FROM db.my_table;
+SELECT id, "type", "format" FROM db.events;
+
+CREATE TABLE db.events (
+    id       INTEGER,
+    "type"   VARCHAR(30),   -- reserved word — must quote
+    label    VARCHAR(100)
+) PRIMARY INDEX (id);
 ```
 
-In table operator string arguments (`ACCUMULATE`, `IDColumn`, `TargetColumns`, `Accumulate`, etc.), the double-quotes must be embedded **inside** the single-quoted string:
+> **Quoted identifiers are case-sensitive.** Use consistent casing — `"type"` and `"TYPE"` are different identifiers.
+
+### Teradata-Specific Reserved Words — Common Identifier Conflicts
+
+The ANSI SQL reserved words are well-known. The words below are **Teradata-only** — not in the ANSI SQL-99 standard — so agents may not recognize them as reserved. Always quote these when using them as column, table, or alias names.
+
+**Words that frequently appear as column or table names:**
+
+| Reserved word | Commonly appears as | TD since |
+|--------------|---------------------|----------|
+| `TYPE` | transaction type, event type, record type | V2R3 |
+| `FORMAT` | file format, output format, date format | V2R3 |
+| `TITLE` | document title; also controls the TD column display header | V2R3 |
+| `MODE` | processing mode, run mode, lock mode | V2R3 |
+| `ACCOUNT` | account_id, account-related tables | V2R3 |
+| `LOG` | log tables, audit logs, log level | V2R3 |
+| `LOCK` | lock status, concurrency tables | V2R3 |
+| `HASH` | hash keys, checksums, deduplication columns | V2R3 |
+| `REQUEST` | request_id, API and service event tables | V2R3 |
+| `STATISTICS` | monitoring tables, collected stats columns | V2R3 |
+| `JOURNAL` | financial journals, transaction audit logs | V2R3 |
+| `CLUSTER` | cluster_id, partition or segment label | V2R3 |
+| `NAMED` | TD column alias syntax (`expr (NAMED alias)`) — risky as a column name | V2R3 |
+| `ENABLED` / `DISABLED` | feature flag columns, configuration status | V2R3 |
+| `CLASS` | object class, classification, CSS class | V2R5 |
+| `PROFILE` | user profiles, configuration profiles | V2R5 |
+| `SUMMARY` | summary text columns, reporting tables | V2R5 |
+| `THRESHOLD` | alert thresholds, monitoring limit columns | V2R5 |
+| `TRACE` | trace_id, debug or telemetry columns | V2R5 |
+
+**Teradata SQL extension keywords — these are clause keywords, not identifiers:**
+
+| Keyword | Purpose |
+|---------|---------|
+| `QUALIFY` | Filters window function results — like WHERE for OVER clauses; not in ANSI SQL |
+| `SAMPLE` | Random row sampling: `SELECT * FROM t SAMPLE 100` or `SAMPLE .05` |
+| `VOLATILE` | Session-scoped temp table: `CREATE VOLATILE TABLE ...` |
+| `LOCKING` | Lock modifier: `LOCKING TABLE t FOR ACCESS SELECT ...` |
+| `REPLACE` | TD DDL: `REPLACE VIEW` — not `CREATE OR REPLACE` |
+| `EXPLAIN` | Execution plan: `EXPLAIN SELECT ...` |
+| `FALLBACK` | Table-level data protection option at `CREATE TABLE` time |
+| `MULTISET` | Table type allowing duplicate rows: `CREATE MULTISET TABLE ...` |
+| `MACRO` | Stored parameterized query: `CREATE MACRO ...` |
+| `COLLECT` | Statistics collection: `COLLECT STATISTICS ON db.t COLUMN (col)` |
+| `BT` / `ET` | Begin Transaction / End Transaction |
+| `SEL` | Teradata shorthand for `SELECT` |
+
+### Reserved Words in Table Operator String Arguments
+
+In table operator clauses that take column names as **string arguments** (`Accumulate`, `IDColumn`, `TargetColumns`, `ResponseColumn`, etc.), double-quotes must be embedded **inside** the single-quoted string:
 
 ```sql
 -- WRONG: 'type' is a reserved word — Teradata will reject or misparse this
@@ -158,19 +213,16 @@ USING IDColumn('id') Accumulate('type', 'value')
 USING IDColumn('id') Accumulate('"type"', '"value"')
 ```
 
-This applies to **any** table operator clause that takes column names as string arguments:
+This applies to any table operator clause that takes column names as string arguments:
 
 ```sql
--- All of these follow the same rule
 IDColumn('"type"')
 TargetColumns('"value"', '"date"', 'non_reserved_col')
 Accumulate('"type"', 'amount', '"date"')
 ResponseColumn('"value"')
 ```
 
-**When in doubt, quote it.** Double-quoting a non-reserved word in a string argument is harmless; leaving a reserved word unquoted will cause a parse error.
-
-Common Teradata reserved words that appear as column names: `type`, `date`, `time`, `timestamp`, `value`, `name`, `format`, `title`, `level`, `mode`, `status`, `class`, `key`, `index`, `year`, `month`, `day`, `hour`, `minute`, `second`.
+**When in doubt, quote it.** Double-quoting a non-reserved word inside a string argument is harmless; leaving a reserved word unquoted will cause a parse error.
 
 ## Teradata Operator Differences