Skip to main content

juuri

Juuri (pronounced "JooRi", Finnish for root): Knowledge Graphs for RAG. Nothing else.

pip install juuri

Status

v0.3: the knowledge-graph RAG loop works end to end, with temporal facts, root-cause tracing, grounding checks, entity resolution, community summaries, text-to-query, retrieval evaluation, conflict detection and exports. The API is early and may change.

Step What you get
Store Typed (subject, predicate, object) triples, integer-interned strings, S/P/O indexes
Schema Allowed classes, relations with domain/range, subclasses, rules, constraints
Extract LLM triple extraction through Ollama, including dates when the text gives them
Provenance Source, chunk ID, confidence and timestamp on every fact
Temporal valid_from / valid_to on facts, "as of" queries, fact history
Re-ingest Re-ingesting a named source replaces only that source's facts
Resolve Merge names for the same entity (sameAs), with suggestions
Retrieve Vector search for seed entities, then k-hop expansion
Traverse Paths between entities, root-cause and blast-radius chains
Communities Groups of linked entities with cited LLM summaries, for broad questions
Ask Text-to-query: the LLM writes a structured query, juuri validates and runs it
Context Token-budgeted prompt text with citations
Verify Grounding check of an LLM answer against retrieved facts
Evaluate Hit rate, coverage, precision and MRR against gold answers, plus settings tuning
Conflicts Facts that contradict each other under schema constraints, with dates considered
Export Turtle / OWL, and a one-way networkx graph

Quick start

Needs a running Ollama server with a chat model and an embedding model, for example:

ollama pull qwen2.5:7b-instruct
ollama pull nomic-embed-text
import juuri as jr

schema = jr.Schema(
    classes=["Player", "Team"],
    relations={
        "PLAYED_FOR": ("Player", "Team"),
        "CAPTAINED":  ("Player", "Team"),
    },
)

store = jr.TripleStore(schema=schema)

# text -> triples (each fact keeps its source, chunk, confidence, timestamp)
llm = jr.OllamaLLM("qwen2.5:7b-instruct")
store.ingest({"cricket.txt": "MS Dhoni captained Chennai Super Kings."}, llm=llm)

# embed entities, then retrieve
embedder = jr.OllamaEmbedder("nomic-embed-text")
store.embed(embedder)
hits = store.retrieve("Who led CSK?", embedder, seeds=1, hops=1)

# cited prompt text for your LLM
print(hits.to_context(tokens=2000))
FACTS:
- MS Dhoni CAPTAINED Chennai Super Kings [cricket.txt#0]

SOURCES:
[cricket.txt#0] MS Dhoni captained Chennai Super Kings.

Using the pieces

Triples and lookups

store = jr.TripleStore()
store.add("Dhoni", "PLAYED_FOR", "CSK", source="stats.csv", confidence=0.9)
list(store.find(obj="CSK"))                 # None is a wildcard
store.provenance("Dhoni", "PLAYED_FOR", "CSK")

Adding the same fact again from another source keeps one triple and adds a second provenance record.

Schema guard

With a schema, add() rejects undeclared relations and wrong subject or object types by raising jr.SchemaError. Pass types the first time an entity appears: store.add("Dhoni", "PLAYED_FOR", "CSK", "Player", "Team"). The type is remembered afterwards. Subclasses count as their parent class (subclasses={"Batsman": "Player"}).

Rules (transitive, symmetric, implies:REL, inverse:REL) are validated and stored. verify uses symmetric, inverse and implies, and all of them are written to the Turtle export. transitive is stored and exported but not applied anywhere.

Constraints (functional, inverse_functional, exclusive:REL) are a separate argument, constraints=, used by conflicts() (see below). add() does not reject facts that break them.

Ingest

store.ingest(docs, llm, max_chars=2000, min_confidence=0.0, source=None) accepts a string, a list of strings, or a dict {source_name: text}. Triples that break the schema are dropped. If the text says when a fact was true, the LLM adds valid_from / valid_to. A triple with an unreadable date is skipped, not guessed. Output that is not valid JSON raises jr.ExtractionError. A failed Ollama call raises jr.LLMError. Any object with complete(prompt) -> str works as an LLM backend.

Retrieve

store.retrieve(query, embedder, seeds=3, hops=1, expand=None, direction="both", min_score=0.0)

  • expand=["CAUSED_BY"] follows only those relations.
  • direction is "out", "in" or "both".
  • Call store.embed(embedder) first, and again after adding facts about new entities.
  • Any object with embed(list_of_texts) -> list_of_vectors works as an embedder.

Context

hits.to_context(tokens=2000, passages=True, count_tokens=None) adds facts in rank order until the budget is used, then source passages with what is left. The default token count is a rough estimate of 1 token per 4 characters. Pass count_tokens with a real tokenizer for exact budgets.

Re-ingest

ingest with a source name replaces that source's earlier facts (replace=True is the default), so an edited document leaves no stale facts. A fact that another source also backs stays. All documents are extracted before anything is replaced, so a failed LLM call changes nothing. store.sources() lists sources and store.remove_source(name) deletes one. Facts added without a source are never removed, and a document ingested without a source name cannot be replaced.

Traverse

store.root_cause("A")            # follow CAUSED_BY to the end of each branch
store.blast_radius("D")          # follow it backwards: what D led to
store.paths("A", "D", max_hops=4, predicates=["CAUSED_BY"])

for chain in store.root_cause("A"):
    print(chain.to_context("ROOT CAUSE CHAIN"))   # numbered, cited lines

Direction follows how triples are stored: A CAUSED_BY B means B is a cause of A, so root_cause walks subject to object. Each Chain has .start, .end, .triples, .provenance, .cyclic and .truncated. A cycle stops the walk with cyclic=True and means there is no single root cause.

Verify

v = store.verify(answer, hits, llm)   # needs a schema
v.grounded                            # True only if every claim is supported
v.score                               # share of supported claims, or None if no claims
v.supported, v.unsupported, v.unchecked

The LLM lists the answer's claims and each one is matched against the retrieved facts, using the schema's symmetric, inverse and implies rules. unchecked means the claim uses a relation the schema does not declare. Invalid LLM output raises jr.ExtractionError. A true claim that retrieval did not return counts as unsupported.

Entity resolution

store.suggest_merges()                 # candidates only, nothing is merged
store.merge("Dhoni", "MS Dhoni")       # alias -> canonical
store.resolve_entities()               # merges names identical after normalising
store.aliases("MS Dhoni")

After a merge, the alias name works in add, find, paths, root_cause, blast_radius and verify. Merging entities of different types raises SchemaError. resolve_entities() defaults to min_score=1.0 on purpose: "Dhoni" and "MS Dhoni" score 0.9, and short names can match several different entities, so review those with suggest_merges() first. Pass min_embedding to suggest_merges to add suggestions from stored vectors.

Temporal facts

store.add("Dhoni", "CAPTAINED", "CSK", valid_from="2008", valid_to="2015")
store.add("Dhoni", "CAPTAINED", "RPS", valid_from="2016", valid_to="2017")

list(store.find("Dhoni", "CAPTAINED", as_of="2016-06-01"))   # RPS only
store.history("Dhoni")                                       # each fact with its periods
hits = store.retrieve("Who did Dhoni captain?", embedder, as_of="2016-06-01")
  • Dates are "YYYY", "YYYY-MM", "YYYY-MM-DD", an int year or a date. A partial valid_from means the start of the period and a partial valid_to means its end, so valid_to="2015" is 2015-12-31. Both ends are inclusive and None is open.
  • The dates live on the provenance record, so each source can give its own period. The same fact from two sources with different periods stays one triple with both periods.
  • as_of must be a full date. A fact with dated evidence is valid if one of its periods contains the day. A fact with no dated evidence is kept unless you pass include_undated=False.
  • as_of works in find, retrieve, paths, root_cause, blast_radius, communities, conflicts, query and both exports.
  • to_context() prints the periods, for example (valid 2008-01-01 to 2015-12-31).
  • This stores periods only. A newer fact does not close an earlier period automatically.

Communities and broad questions

store.communities(min_size=2, resolution=1.0)            # groups of linked entities
store.summarize_communities(llm)                         # a cited summary for each
print(store.global_context("Tell me about the Bangalore team", embedder, llm, k=3, tokens=2000))

Communities come from a pure-Python Louvain algorithm, with any group that is not connected inside itself split afterwards. It is not Leiden, and there is one flat level, not a hierarchy. Direction and relation names are ignored, and several facts between the same pair add weight. Summaries are cached by the exact facts they cover, so an unchanged community is not sent to the LLM again. A community larger than max_facts (default 60) is summarised from its first facts only, and global_context says so.

Ask (text-to-query)

result = store.ask("Which team did Dhoni captain in 2016?", llm)   # needs a schema
result.query                  # the validated query that ran, with graph entity names
result.rows                   # for "match": bindings plus the facts and provenance behind them
result.chains                 # for root_cause, blast_radius and paths
print(result.to_context(tokens=2000))

The LLM writes a small JSON query. juuri checks it against the schema, links entity names to real graph entities, and runs it. The LLM never writes code. store.query(spec) runs the same JSON directly. There are four operations:

{"op": "match", "patterns": [{"subject": "Dhoni", "predicate": "CAPTAINED", "object": "?team"},
                             {"subject": "?player", "predicate": "PLAYED_FOR", "object": "?team"}],
 "as_of": "2010-05-01", "limit": 50}
{"op": "root_cause",   "entity": "Login timeout", "relation": "CAUSED_BY"}
{"op": "blast_radius", "entity": "Disk full",     "relation": "CAUSED_BY"}
{"op": "paths", "source": "Login timeout", "target": "Disk full", "predicates": ["CAUSED_BY"]}

A term starting with ? is a variable. Unknown keys, undeclared relations, relation variables and partial dates raise jr.QueryError. A name that matches no entity raises QueryError, and so does a name that matches several, with the candidates listed. It never guesses. There is no OR, NOT or counting. Check the output of ask for your own model: a wrong query that is still valid runs and returns wrong rows.

Evaluate retrieval

cases = [{"question": "Who led CSK?",
          "triples": [["MS Dhoni", "CAPTAINED", "Chennai Super Kings"]],
          "as_of": "2010-05-01"}]                          # as_of is optional
ev = store.evaluate(embedder, cases, seeds=2, hops=1)
print(ev.summary())
for run in store.tune(embedder, cases, {"hops": [0, 1, 2], "seeds": [1, 3]}):   # best first
    print(run.params, run.coverage, run.precision)

Metrics are hit rate, coverage (recall), precision, mean reciprocal rank, and context coverage (how many gold facts survive the to_context token budget). Gold facts that are not in the graph at all are counted separately, so an ingestion miss is not blamed on retrieval. Precision is low by design, because retrieval returns neighbouring facts. tune sorts by coverage, then hit rate, then fewest context tokens. With few cases many settings tie, so read the other columns before trusting the top row.

Conflicts

schema = jr.Schema(
    classes=["Person", "Team", "City"],
    relations={"BORN_IN": ("Person", "City"), "CAPTAINED": ("Person", "Team"),
               "MANAGES": ("Person", "Team"), "COACHES": ("Person", "Team")},
    constraints={"BORN_IN": "functional",              # one object per subject
                 "CAPTAINED": "inverse_functional",    # one subject per object
                 "MANAGES": ["exclusive:COACHES"]},    # cannot both hold for the same pair
)
for conflict in store.conflicts():     # also: as_of, include_undated, relations, cross_source_only, limit
    print(conflict.to_context())

Two facts conflict only if their validity periods overlap, so captaining one team until 2015 and another from 2016 is fine. If either fact has no dated evidence, the conflict is reported with undated=True, meaning the facts may not overlap. Each side carries its provenance, and conflict.cross_source is true when the two sides share no source. juuri reports conflicts but does not decide which side is right. Entity names are compared as stored, so merge variants such as Chennai and Chennai, India first.

Export

store.to_turtle("graph.ttl", base="https://data.acme.com/kg/")   # OWL 2 vocabulary, loads in Protégé
g = store.to_networkx()          # networkx.MultiDiGraph, needs: pip install juuri[networkx]

The Turtle file contains the schema as an ontology (classes, subclasses, relations with domain and range, transitive, symmetric, inverse and implies rules, functional constraints), the entities, merged aliases as owl:sameAs, and the facts. With include_provenance=True (the default) each evidence record becomes an rdf:Statement with its source, chunk, confidence, dates and recording time, which you can query in SPARQL. The output is deterministic. There is no Turtle import. In the networkx graph each fact is an edge keyed by its predicate, with sources, chunks, confidence, periods and provenance attributes. Both exports are copies, and both accept as_of.

Limitations

  • Everything is in memory. Nothing is saved to disk yet, including aliases, vectors and summaries.
  • Vector search compares the query with every entity. There is no ANN index.
  • Only entities are embedded, not edges.
  • Merges and source removals rebuild the store, and suggest_merges compares every pair within a type, so they are slow on large graphs.
  • verify applies rules one step at a time. There is no rule chaining, and it ignores time, so retrieve with as_of first.
  • Constraints are checked on demand by conflicts(), and exported to OWL, but never enforced by add().
  • Community detection is Louvain, not Leiden, and untested on large graphs.
  • Ollama is the only built-in backend.

Roadmap

See juuri-base.md for the design notes and the v0.1 to v0.3 roadmap.

Metadata

Release files for juuri 0.3.0

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for juuri 0.3.0
File Size Uploaded Provenance Actions
juuri-0.3.0.tar.gz 48.3 kB Without provenance Details for juuri-0.3.0.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for juuri 0.3.0
File Interpreter ABI Platform Actions
juuri-0.3.0-py3-none-any.whl Python 3 none any Details for juuri-0.3.0-py3-none-any.whl

Total release size: 98.4 kB

Release files / juuri-0.3.0.tar.gz

Download URL juuri-0.3.0.tar.gz
Size 48.3 kB
Tags Source
SHA-256 checksum
How to use checksums
1912dbe9ee13272daa7d294e3be87295c5eebc6d5f8f1be88a153b4a78238e5e
BLAKE2b-256 checksum
How to use checksums
cf878c2e1f2cfb33d7b45a0397e0f89132adf9c5e5594302431e9b1892aa9e1c
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/6.2.0 CPython/3.12.12

Release files / juuri-0.3.0-py3-none-any.whl

Download URL juuri-0.3.0-py3-none-any.whl
Size 50.1 kB
Tags Python 3
SHA-256 checksum
How to use checksums
ed13d85cee6dee3a3c170d96d59d94217404d3b27e7409cb57b4b9ad6c830fa3
BLAKE2b-256 checksum
How to use checksums
fbf6fabcd450344786cf089394389d219f0c494d2031614a21b472a6bcb721e7
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/6.2.0 CPython/3.12.12

Security for juuri 0.3.0

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page